MM-FinEval: A Multi-Task Multimodal Benchmark for Real-World Financial Forecasting

# A Multi-Task Multimodal Benchmarkfor Real-World Financial Forecasting
Source: [https://arxiv.org/html/2609.38523](https://arxiv.org/html/2609.38523)
## MM\-FinEval: A Multi\-Task Multimodal Benchmark for Real\-World Financial Forecasting

Yanguang LiuAffiliation:New Jersey Institute of TechnologyHuopu ZhangAffiliation:Georgia Institute of TechnologySaisai HuAffiliation:Pace UniversityHaiyan ZhaoAffiliation:New Jersey Institute of TechnologyHekun HuangAffiliation:New Jersey Institute of TechnologyMengnan DuAffiliation:The Chinese University of Hong Kong, Shenzhen

###### Abstract

Financial forecasting from earnings conference calls requires models to reason over complex corporate disclosures, market expectations, and subtle communication signals\. However, existing financial benchmarks are often limited to unimodal inputs or single\-task settings, making it difficult to evaluate whether multimodal large language models \(LLMs\) can support real\-world financial analysis\. In this paper, we introduce MM\-FinEval, a novel benchmark designed to evaluate multimodal LLMs across multiple financial tasks\. MM\-FinEval spans a diverse timeline from 2019 to 2022\. The entire proposed dataset contains 2,045 S&P 500 conference earning calls as inputs and 12 financial task labels as outputs\. Each input contains three modalities: a word\-to\-word text transcript of the earning call, the corresponding presentation slides used during the call, and the entire audio recording\. To establish a rigorous evaluation framework, we analyze 19 baseline models across three distinct model categories: Image\-Text, Audio\-Text, and Any\-to\-Any configurations\. We observe that small\-size Any\-to\-Any models processing all three modalities achieve strong performance, even when compared against larger proprietary models restricted to two\-modality inputs\. This indicates that our tri\-modal dataset design introduces useful, non\-redundant information\. These results validate that text, audio, and visual data serve as important, complementary signals that mimic the decision\-making process of expert human analysts\. The MM\-FinEval dataset is available at[https://github\.com/Tizzzzy/MM\-FinEval](https://github.com/Tizzzzy/MM-FinEval)\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2609.38523v1/pipeline.png)Figure 1:MM\-FinEval Construction Pipeline\. The dataset construction follows a three\-stage process: assembling a foundational collection of publicly traded US firms, systematically web\-scraping synchronized tri\-modal data \(text transcripts, audio recordings, and slides\), and annotating the multimodal inputs with 12 financial task labels\.Earnings conference calls are important events for public companies, serving as a primary channel for management to communicate financial performance, strategic outlooks, and operational updates to the market\. The quality of a company’s earnings conference call is directly linked to the next\-day stock market performance\. In practice, financial analysts rely heavily on these events, extracting signals from company presentations and Q&A sessions to analyze corporate health and forecast future performance\. Concurrently, there is a rapidly growing trend of leveraging LLMs to analyze financial markets and automate data processing[Wu et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib1);[Li et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib2);[Huang et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib3);[Yang et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib4)\. If LLMs can be trained to successfully capture the dense, multi\-layered information presented during these calls, financial institutions could benefit from an automated and scalable AI finance analysis pipeline that has historically required substantial human analysis\.

Despite the rapid advancement of LLMs, existing financial models remain constrained by their reliance on uni\-modal architectures and their tendency to focus on narrow, specific financial tasks[He and Zhang \(2025\)](https://arxiv.org/html/2609.38523#bib.bib5);[Zhu et al\. \(2025\)](https://arxiv.org/html/2609.38523#bib.bib6);[Liang and Carrasco Kind \(2025\)](https://arxiv.org/html/2609.38523#bib.bib7);[Konstantinidis et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib8);[Liu et al\. \(2025b\)](https://arxiv.org/html/2609.38523#bib.bib9)\. A major reason for this limitation is that the majority of current financial datasets are heavily skewed toward a single modality, such as pure text or numerical time\-series data[Chen et al\. \(2021\)](https://arxiv.org/html/2609.38523#bib.bib11);[Lai et al\. \(2025\)](https://arxiv.org/html/2609.38523#bib.bib12);[Shah et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib13);[Dong et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib14)\. Furthermore, even when recent datasets incorporate multimodal elements beyond text, they are typically restricted to a specific task[Shu et al\. \(2025a\)](https://arxiv.org/html/2609.38523#bib.bib10);[Li et al\. \(2020\)](https://arxiv.org/html/2609.38523#bib.bib17);[Luo et al\. \(2025\)](https://arxiv.org/html/2609.38523#bib.bib16);[Shu et al\. \(2025b\)](https://arxiv.org/html/2609.38523#bib.bib15)\. This combination of unimodal and task\-specific training data severely restricts financial LLMs to learn from additional contextual information, ultimately limiting their broader applicability in complex, real\-world financial environments\.

To bridge this gap, we propose MM\-FinEval, a large\-scale, multi\-modal, and multi\-label financial benchmark dataset designed specifically for earnings call analysis\. Spanning a period from 2019 to 2022, the dataset encompasses a total of 2045 data points from S&P 500 companies, with 450 from 2019, 273 from 2020, 499 from 2021, and 823 from 2022\. Each data point integrates three modalities: word\-to\-word text transcripts of the earnings call, the corresponding raw earnings call audio, and the accompanying presentation slide decks\. Furthermore, MM\-FinEval is annotated with 12 labels, providing a benchmark to evaluate the performance of multimodal financial models\. Through an extensive experiment with 19 state\-of\-the\-art models, we found that:

- •Current LLMs exhibit a severe performance gap between market return and volatility forecasting\.
- •Integrating all three modalities provides performance advantage, proving that non\-redundant visual and audio information are essential for replicating expert financial reasoning\.
- •Model performance on volatility tasks aligns with the financial theory of information half\-life\. As the forecasting horizon expands from 3 to 30 days, the earnings signal is rapidly diluted by unpredictable market factors\.

## 2Related Work

### 2\.1Earnings Conference Call Datasets

Earnings conference calls are important events for public companies because the quality of the information disclosed during these presentations and Q&A sessions is directly linked to next\-day stock market performance\. Recently, researchers have increasingly deployed machine learning models to systematically interpret the information in these calls[Sawhney et al\. \(2020b\)](https://arxiv.org/html/2609.38523#bib.bib19);[Sawhney et al\. \(2020a\)](https://arxiv.org/html/2609.38523#bib.bib20);[Mathur et al\. \(2022\)](https://arxiv.org/html/2609.38523#bib.bib21);[Sawhney et al\. \(2020c\)](https://arxiv.org/html/2609.38523#bib.bib22), aiming to uncover quantitative investment signals and making money\. This effort to uncover quantitative investment signals has motivated the community to publish specialized datasets focusing on earnings conference calls\. For example, ECTSum[Mukherjee et al\. \(2022\)](https://arxiv.org/html/2609.38523#bib.bib18)provides benchmarks for bullet\-point summarization of highly verbose transcripts\. FinNLI[Magomere et al\. \(2025\)](https://arxiv.org/html/2609.38523#bib.bib23)is developed to assess logical deduction\. SubjECTive\-QA[Pardawala et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib24)measures the subjectivity of Q&A responses across conference earning calls\. Recent large text datasets like STRUX[Lu et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib25)and DEC[Yu et al\. \(2025\)](https://arxiv.org/html/2609.38523#bib.bib26)created for LLM\-driven investment decision\-making\.

However, these existing datasets remain constrained by their reliance on unimodal architectures, do not include the acoustic and visual signals inherent in the actual conference call\. EC[Qin and Yang \(2019\)](https://arxiv.org/html/2609.38523#bib.bib27)which was the first dataset to combine textual transcripts with raw audio recordings to predict financial risk\. Building upon this, MAEC[Li et al\. \(2020\)](https://arxiv.org/html/2609.38523#bib.bib17)scaled the available conference call data and optimized to predict stock volatility\. Datsets such as Earnings\-21[Del Rio et al\. \(2021\)](https://arxiv.org/html/2609.38523#bib.bib28)and ConEC[Huang et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib29)focus heavily on the acoustic challenges of the domain, serving primarily as benchmarks for automatic speech recognition rather than broader financial reasoning\. FinCall\-Surprise[Shu et al\. \(2025a\)](https://arxiv.org/html/2609.38523#bib.bib10), which provides conference call text transcripts, audio, and slides, is optimized exclusively for predicting earnings surprises\. Despite the success of these existing multi\-modal conference call datasets, they are limited to a specific financial task\. This severely restricts the ability of financial LLMs to learn from additional contextual information, directly motivating the need for multi\-label multi\-modal benchmarks\.

Table 1:Data statistics of the MM\-FinEval\. The dataset spans four years \(2019–2022\), with each conference call containing three synchronized modalities: text transcripts, presentation slides, and audio recordings\. We report the minimum, maximum, and average values for each modality\. Transcript length is measured in words, slides in pages, and audio in seconds\. The last column shows how much data we have for each year\.

## 3MM\-FinEval Construction

The construction of MM\-FinEval dataset follows a three\-stage pipeline designed to collect, synchronize, and annotate data from multiple sources\. As illustrated in Figure[1](https://arxiv.org/html/2609.38523#S1.F1), the first stage involves gathering a set of liquid public firms in the United States\. To ensure data quality, we restrict the sample to companies with a market capitalization exceeding $1 billion and an average daily trading volume above $50 million\. This initial screening process yielded a pool of over 4,000 companies\.

In the second stage, we acquire data across the three modalities for each firm’s earnings conference calls\. We systematically web\-scrape these contents from official company websites\. The collected text transcripts are highly structured, featuring clear speaker identifiers \(operators, executives, and analysts\) alongside their corresponding statements\. Because most publicly traded firms host one earnings call per fiscal quarter, we ensure temporal consistency by manually aligning all data sources according to their specific reporting dates\. Furthermore, we cross\-reference the conference call titles to verify that the transcript, audio recording, and presentation slides all originate from the exact same event, ensuring synchronization across our multimodal dataset\.

### 3\.1Multi\-Label Preliminaries

MM\-FinEval benchmark includes 12 labels designed to predict a company’s cumulative abnormal return, stock volatility, and corporate bond returns based on its quarterly earnings call\. The complete labels are shown on the right side of Figure[1](https://arxiv.org/html/2609.38523#S1.F1)\.

#### Cumulative Abnormal Return\.

Cumulative abnormal return \(CAR\) measures the market\-adjusted price reaction around an earnings call event[Hirshleifer et al\. \(2009\)](https://arxiv.org/html/2609.38523#bib.bib34);[Duan et al\. \(2018\)](https://arxiv.org/html/2609.38523#bib.bib35)\. For each firmiiand event datett, we first compute the abnormal return by subtracting a benchmark market return from the firm’s realized return:

A​Ri,t=Ri,t−Rm,t,AR\_\{i,t\}=R\_\{i,t\}\-R\_\{m,t\},\(1\)whereRi,tR\_\{i,t\}denotes the stock return of firmiion daytt, andRm,tR\_\{m,t\}denotes the corresponding market return\. We then aggregate abnormal returns over an event window\[t1,t2\]\[t\_\{1\},t\_\{2\}\]:

C​A​Ri,\[t1,t2\]=∑t=t1t2A​Ri,t\.CAR\_\{i,\[t\_\{1\},t\_\{2\}\]\}=\\sum\_\{t=t\_\{1\}\}^\{t\_\{2\}\}AR\_\{i,t\}\.\(2\)CAR reflects whether the information revealed in the earnings call leads to a positive or negative market reassessment of the firm\. We include CAR\-based labels over multiple post\-call windows to evaluate whether models can infer short\-term and medium\-term directional market reactions from multimodal earnings\-call information\.

#### Volatility\.

Volatility measures stock\-price fluctuations over a finite period and is commonly used as a measure of risk[Kogan et al\. \(2009\)](https://arxiv.org/html/2609.38523#bib.bib36)\. In our study, the period is defined relative to the earnings conference call using pre\-event and post\-event windows of different lengths\. Following Kogan et al\.[Kogan et al\. \(2009\)](https://arxiv.org/html/2609.38523#bib.bib36), volatility from dayt−τt\-\\tauto dayttis defined as

v\[t−τ,t\]=∑i=0τ\(rt−i−r¯\)2τ,v\_\{\[t\-\\tau,t\]\}=\\sqrt\{\\frac\{\\sum\_\{i=0\}^\{\\tau\}\(r\_\{t\-i\}\-\\bar\{r\}\)^\{2\}\}\{\\tau\}\},\(3\)
wherertr\_\{t\}is the return on daytt,r¯\\bar\{r\}is the mean return over the period of dayt−τt\-\\tauto daytt\.rt=Pt/Pt−1−1r\_\{t\}=P\_\{t\}/P\_\{t\-1\}\-1is the return price, withPtP\_\{t\}denoting the closing price on daytt\. While CAR captures the direction of the market response, volatility captures stock\-price instability regardless of whether returns are positive or negative\. The labelsfut\_3d–fut\_30dandpast\_3d–past\_30dmeasure post\-event and pre\-event volatility, respectively, over windows of 3, 7, 15, and 30 calendar days\.

#### Corporate Bond Return\.

Corporate bond returns capture the monthly performance of a firm’s outstanding bonds and provide a debt\-market measure of investors’ responses to information disclosed during earnings calls\. The bond return data are observed at the monthly level and are matched to conference\-call observations using the firm’s stock ticker, recorded asCOMPANY\_SYMBOL, and the bond trading month, recorded inDATE\. Specifically, monthly returns are computed as the percentage change in the bond’s dirty price \(clean price plus accrued interest\), adjusted for coupon payments received during the holding period, following the standard methodology in the corporate bond literature[Bessembinder et al\. \(2008\)](https://arxiv.org/html/2609.38523#bib.bib32);[Jostova et al\. \(2013\)](https://arxiv.org/html/2609.38523#bib.bib33)\. To account for differences in bond trading frequency and liquidity, we consider three complementary return measures that differ only in their definition of the month\-end price\.RET\_EOMis calculated using the last observed transaction price during the calendar month, regardless of the trading date\.RET\_LDMuses the transaction price on the last trading day of the month and is missing if the bond does not trade on that day\.RET\_L5Muses the last transaction price observed within the final five trading days of the month, providing a balance between month\-end pricing and sample coverage\.

## 4Data Analysis

Table[1](https://arxiv.org/html/2609.38523#S2.T1)presents a detailed statistical analysis of our dataset, summarizing its scale and multimodal structure\. MM\-FinEval has a total of 2045 data spanning from 2019 to 2022\. On average, an individual data features a text transcript of approximately 9,000 words, accompanied by a 27\-page presentation slide deck and an audio recording spanning roughly 3,700 seconds\. As illustrated in Figure[2](https://arxiv.org/html/2609.38523#S4.F2)a, the content of these calls is consistently structured around three primary speaker roles: Operators, Executives, and Analysts\. Operators typically manage the flow of the event by opening and closing the call\. Executives, such as the CEO and CFO, deliver the company’s financial results and strategic outlook\. Finally, Analysts representing investment firms and financial institutions ask questions to obtain more information into the company’s operational and financial performance\. More data analysis including our label statistic is in Appendix[A](https://arxiv.org/html/2609.38523#A1)\.

Figure 2:Transcript Structure and Label Statistics\.
## 5Benchmark Settings

### 5\.1Baseline LLMs

To establish a comprehensive performance benchmark for MM\-FinEval, we selected 19 state\-of\-the\-art open\- and closed\-source models\. These models were categorized into three architectural groups based on their input capabilities\.

The first group comprises Large Vision\-Language Models \(LVLMs\), which process both image and text inputs\. This group includes the Qwen 3\.5 series \(spanning 2B to 35B parameters\)[Qwen Team \(2026a\)](https://arxiv.org/html/2609.38523#bib.bib37), the Qwen 3\.6 series \(27B and 35B\)[Qwen Team \(2026b\)](https://arxiv.org/html/2609.38523#bib.bib38), Gemma 3n[Team et al\. \(2024\)](https://arxiv.org/html/2609.38523#bib.bib39), and the Mistral 3 series \(3B, 8B, and 14B\)[Jiang et al\. \(2023\)](https://arxiv.org/html/2609.38523#bib.bib40)\. It also includes proprietary models: GPT\-5\.4, GPT\-5\.5, Claude 4\.6 Sonnet, and Claude 4\.8 Opus\. The second group consists of Text\-Audio Language Models, which process both audio and text inputs, represented in our study by Voxtral 24B[Liu et al\. \(2025a\)](https://arxiv.org/html/2609.38523#bib.bib41)and Moss 8B[Team \(2026b\)](https://arxiv.org/html/2609.38523#bib.bib42)\. Finally, the third group features Any\-to\-Any Models, which are natively equipped to process all modalities, text, audio, and images, simultaneously\. This includes Gemma 4 E2B and Gemma 4 E4B[Team \(2026a\)](https://arxiv.org/html/2609.38523#bib.bib43)\.

### 5\.2Benchmark Input Design

To systematically evaluate the performance of the baseline models, we designed a customized input and prompting strategy tailored to each modality\. The full set of prompts used in the experiments is provided in Appendix[B](https://arxiv.org/html/2609.38523#A2)\.

#### Image\-Text Modality:

Because the selected baseline models represent the current state\-of\-the\-art, they all support extended context windows exceeding 128K tokens\. Given that the longest text transcript in our dataset is 21,894 words, we input the raw transcripts directly without applying any preprocessing steps, such as textual summarization\. However, the visual data required structural adaptation\. The original presentation slides are collected in PDF format\. Since most models do not natively support PDF input, we converted each page into a separate image\. To accommodate the image\-count limits inherent to many of these models, we implemented a random sampling strategy\. For each presentation, we randomly selected five images from the slide deck, explicitly excluding the first and last pages, which typically contain non\-substantive content such as title cards or concluding remarks\. In cases where a presentation contained five or fewer substantive slides, all available slides were used\. The final input sequence was structured as: “\{Images\} \{Instruction\} \{Text Transcript\} \{Answer Format\}”\. Although this specific benchmark experiment restricts the visual input to five sampled images, the complete PDF files are included in the public dataset release\.

#### Audio\-Text Modality:

We structured the input sequence as: “\{Audio\} \{Instruction\} \{Text Transcript\} \{Answer Format\}”\.

#### Image\-Audio\-Text Modality:

To ensure a fair comparison across all baseline models, we maintained the exact input constraints established in the previous setups\. We used the same randomly selected slides as the Image\-Text setup\. The final input sequence for this comprehensive modality was structured as: “\{Images\} \{Audio\} \{Instruction\} \{Text Transcript\} \{Answer Format\}”\.

Table 2:Overall Performance of 19 Baseline Models on MM\-FinEval\. The evaluation reports averaged Mean Absolute Error \(MAE\) and Directional Accuracy \(DA\) across 12 financial tasks spanning the 2019–2022 period\. For each row, the highest score for each metric is highlighted in a distinct color corresponding to the model type\.

### 5\.3Benchmark Metrics

Because all of our labels are related to stock returns or volatility, the ground truth data consists of continuous decimal values\. Therefore, we frame the prediction as a regression task\. In this context, Mean Absolute Error \(MAE\) serves as the default metric for evaluating absolute model performance on our benchmark\. For a dataset ofnnsamples, MAE is defined as:

MAE=1n​∑i=1n\|yi−y^i\|,\\text\{MAE\}=\\frac\{1\}\{n\}\\sum\_\{i=1\}^\{n\}\|\{\}y\_\{i\}\-\\hat\{y\}\_\{i\}\|\{\},\(4\)
whereyiy\_\{i\}represents the ground\-truth numerical label andy^i\\hat\{y\}\_\{i\}denotes the model’s prediction\. However, given the inherent complexities of financial forecasting, relying on a single evaluation metric is insufficient to capture a model’s true reasoning capabilities\. Therefore, we also employ the Pearson Correlation Coefficient \(rr\), formulated as:

r=∑i=1n\(yi−y¯\)​\(y^i−y^¯\)∑i=1n\(yi−y¯\)2​∑i=1n\(y^i−y^¯\)2\.r=\\frac\{\\sum\_\{i=1\}^\{n\}\(y\_\{i\}\-\\bar\{y\}\)\(\\hat\{y\}\_\{i\}\-\\bar\{\\hat\{y\}\}\)\}\{\\sqrt\{\\sum\_\{i=1\}^\{n\}\(y\_\{i\}\-\\bar\{y\}\)^\{2\}\\sum\_\{i=1\}^\{n\}\(\\hat\{y\}\_\{i\}\-\\bar\{\\hat\{y\}\}\)^\{2\}\}\}\.\(5\)
This metric evaluates the model’s capacity to correctly identify the relative magnitude of market reactions across a broad cross\-section of companies\. Finally, we report Directional Accuracy \(DA\)\. In financial markets, accurately forecasting the direction of a movement is a important signal for investment strategies\. DA measures the proportion of predictions that correctly anticipate the mathematical sign of the actual outcome, expressed as:

DA=1n​∑i=1n𝟏​\(sgn​\(yi\)=sgn​\(y^i\)\),\\text\{DA\}=\\frac\{1\}\{n\}\\sum\_\{i=1\}^\{n\}\\mathbf\{1\}\(\\text\{sgn\}\(y\_\{i\}\)=\\text\{sgn\}\(\\hat\{y\}\_\{i\}\)\),\(6\)
where𝟏\\mathbf\{1\}is the indicator function andsgnextracts the positive or negative sign of the value\.

### 5\.4Implementation Details

All experiments were conducted on two NVIDIA A100 GPUs, each with 80GB of memory\. For all baseline models, we maintained their official repository configurations and loaded them with bfloat16 precision to optimize computational efficiency\. Beyond this precision setting, no other modifications were made to the models’ default parameters, ensuring a fair and reproducible comparison across all baseline evaluations\.

## 6Benchmark Results

### 6\.1The Gap Between Returns and Volatility

As shown in Table[2](https://arxiv.org/html/2609.38523#S5.T2), an initial analysis of the benchmark results reveals a large gap in the ability of LLMs to predict market returns versus market volatility\. Across the majority of baseline models, Directional Accuracy \(DA\) for return\-based tasks \(car, ret\_eom, ret\_l5m, ret\_ldm\) is relatively poor, frequently ranging between 40% and 60%\. Only few models achieve above 70% \(e\.g\., Claude\-4\.8 on ret\_eom\)\. Because DA evaluates only the binary sign of the prediction, a score hovering near 50% indicates that these models only marginally outperform random guessing\. In contrast, the majority of models achieve a DA over 90%, and occasionally reach 100%, when predicting volatility\. Exceptions are rare, with only a few small\-size models, such as Qwen3\.5\-2B, falling below 50%\. However, as illustrated by the label statistics in Figure[2](https://arxiv.org/html/2609.38523#S4.F2)b, it is common knowledge in the financial domain that volatility cannot be negative\. Therefore, despite achieving high DA scores on volatility\-based tasks, small\-size LLMs still fail to preserve a basic understanding of fundamental constraints in financial domains\.

This performance gap is further illuminated by evaluating the Mean Absolute Error \(MAE\)\. For instance, high\-performing models \(e\.g\., Claude, Gemma\-4\) achieve MAEs between 0\.041 and 0\.045 on the Cumulative Abnormal Return \(car\) task\. As shown in Figure[2](https://arxiv.org/html/2609.38523#S4.F2)b, return\-based tasks are inherently noisy: their ground truth means hover near zero while exhibiting high variance\. Consequently, this MAE is large compared to car ground truth mean of 0\.003443 and representing approximately 77% of the dataset’s standard deviation\. In contrast, volatility tasks demonstrate a tighter error distribution\. On tasks like fut\_15d, top models achieve MAEs as low as 0\.0085 to 0\.0097, which represent a much smaller fraction of the ground truth mean \(0\.019874\) and standard deviation \(0\.011657\)\. We present a detailed breakdown of each model’s annual MAE performance in Table[3](https://arxiv.org/html/2609.38523#A3.T3), capturing the granular trends over each consecutive year from 2019\-2022\.

![Refer to caption](https://arxiv.org/html/2609.38523v1/pearson.png)Figure 3:Overall Pearson Correlation Coefficients of Baseline Models\. The heatmap illustrates the models’ capacity to recognize linear trends in financial data\.
### 6\.2Inconsistent Scaling Laws and the Proprietary Advantage

Within specific model architectures and tasks, our benchmark in Table[2](https://arxiv.org/html/2609.38523#S5.T2)reveals a clear scaling law where increased parameter count directly correlates with enhanced financial reasoning capabilities\. This trend is particularly pronounced within the Qwen 3\.5 and 3\.6 model family\. As the parameter size scales from 2 billion to 35 billion, we observe a consistent, upward trajectory in overall Directional Accuracy across both return\-based and volatility tasks\.

However, this scaling law breaks down on the MAE metric\. Rather than exhibiting a consistent decrease as parameter counts grow, the models’ MAE changes unpredictably across tasks\. For example, within the Qwen 3\.5 family on the Cumulative Abnormal Return task, the 27B model yields an MAE of 0\.0534, performing worse than its much smaller 4B \(0\.0498\) and 9B \(0\.0499\) counterparts\. Similar inconsistencies appear across other model families\. For instance, the Mistral\-3B model achieves a lower MAE \(0\.0261\) on the fut\_3d task than both the Mistral\-8B \(0\.0319\) and Mistral\-14B \(0\.0308\)\. These findings suggest that while larger models have a better ability to predict the market movement direction, they do not have the knowledge to accurately estimate the magnitude of those movements\.

Beyond model family scaling, the benchmark results highlight a performance gap between open\-source and proprietary models\. Within the Image\-Text modality, the GPT and Claude model families consistently outperform their open\-source counterparts\. Leading proprietary models, such as GPT\-5\.5 and Claude\-4\.8, achieve the highest tier of Directional Accuracy across the majority of tasks\. Specifically, they reach DA scores of approximately 72% on ret\_eom, and 67% on both ret\_l5m and ret\_ldm\. These results outperform the near\-random guessing baseline \(approximately 50%\) in return based tasks\. Furthermore, their 100% DA on volatility tasks suggests that these advanced proprietary models have a basic understanding of the non\-negativity constraints inherent to financial volatility\.

![Refer to caption](https://arxiv.org/html/2609.38523v1/pearson_trend_vol.png)Figure 4:Annual Pearson Correlation Trends \(2019–2022\) on Volatility Tasks\. The heatmap illustrates how models’ ability to capture linear relationships shifts over time\.
### 6\.3The Advantage of Any\-to\-Any Multimodal Integration

As shown in Table[2](https://arxiv.org/html/2609.38523#S5.T2), our benchmark results show the advantage of integrating all three data modalities: text, audio, and visual\. The Any\-to\-Any models evaluated in this study, specifically Gemma 4 E2B and Gemma 4 E4B, perform well on both return\-based and volatility tasks\. Despite their small size architectures \(Gemma 4 E2B is a 5B parameter model utilizing only 2B active parameters, while Gemma 4 E4B is an 8B model with 4B active parameters\) both achieve a Directional Accuracy exceeding 60% across ret\_eom, ret\_l5m, and ret\_ldm\. This performance is better than much larger Image\-Text modality baselines, such as the 35B parameter Qwen 3\.6 model, and even proprietary GPT 5\.4 model\. Furthermore, on the Cumulative Abnormal Return task, both Gemma 4 E2B \(0\.0418\) and Gemma 4 E4B \(0\.0437\) achieve lower MAEs than the highly capable Claude\-4\.8 model \(0\.0458\)\.

This unexpected performance strongly validates that textual, visual, and audio data streams provide distinct, complementary financial signals rather than redundant information\. In real\-world financial markets, institutional analysts do not evaluate corporate disclosures solely based on the conference call transcript\. They synthesize what management is presenting via visual slides, the literal lexical content of the transcript, and, critically, how that content is delivered, extracting subtle acoustic cues from vocal tone, pacing, and inflection\. To accurately predict market reactions, LLMs must develop a similarly holistic comprehension\. The superior performance of these Any\-to\-Any models shows that equipping LLMs with this complete spectrum of human communication is essential for replicating expert\-level financial reasoning\.

### 6\.4Conservative Versus Aggressive Prediction in Audio\-Text Models

Another interesting observation in Table[2](https://arxiv.org/html/2609.38523#S5.T2)is that, in Audio\-Text columns, Voxtral consistently achieves a superior MAE across the entire benchmark, whereas Moss performs better on DA\. A closer examination reveals that Voxtral exhibits highly conservative prediction behaviors, frequently generating its output close to zero\. By minimizing the absolute magnitude of its predictions, Voxtral successfully avoids absolute error penalties, resulting in a relatively lower MAE\. However, this zero\-anchored conservatism prevents the model from capturing meaningful market momentum, leading to exceptionally poor DA scores\. Conversely, Moss adopts a highly aggressive predictive strategy\. It frequently generates big numerical prediction, and is more likely to capture the correct directional sign\.

![Refer to caption](https://arxiv.org/html/2609.38523v1/DA_trend.png)Figure 5:Annual Directional Accuracy \(DA\) Trends Across MM\-FinEval \(2019–2022\)\. Each line represents an individual model’s performance over the four\-year period, illustrating how predictive accuracy shifts across different financial tasks\.
### 6\.5Pearson Correlation

In addition to MAE and DA, we evaluated the Pearson correlation coefficient to assess the models’ capacity to recognize linear trends and relative shifts in the financial data\. As illustrated in Figure[3](https://arxiv.org/html/2609.38523#S6.F3), open\-source models generally exhibit low Pearson correlations across all tasks, with the highest scores peaking near 0\.20 on the vol\_fut\_15d task \(achieved by Qwen3\.5 27B and Gemma 3n\)\. Interestingly, proprietary models such as GPT\-5\.4, GPT\-5\.5, and the Claude 4 series achieve surprisingly moderate\-to\-strong positive Pearson correlations on volatility tasks\. For instance, GPT\-5\.4 reaches a correlation of 0\.44 on vol\_fut\_7d, and Claude 4\.8 reaches 0\.42\. This results reveals that proprietary architectures can successfully detect relative shifts in market risk\.

We also observe a universal failure across all architectures to predict Cumulative Abnormal Return \(car\)\. While proprietary models demonstrate strong positive correlations on discrete return metrics, with GPT\-5\.5 achieving an impressive 0\.60 on ret\_eom \(End of Month Return\)\. Both open\-source and closed\-source models exhibit near\-zero or even negative correlations on the CAR task\. This discrepancy likely stems from the structural complexity of Cumulative Abnormal Return, which isolates the firm\-specific stock reaction by mathematically regressing against broader market benchmarks over a specific event window\.

Finally, the correlation data exposes a clear temporal degradation in model performance across volatility prediction horizons\. For the proprietary models, predictive capability is highest at the short\-term 3\-day and 7\-day intervals, but gradually decays through the 15\-day mark, ultimately collapsing at the 30\-day horizon for both post\-event and pre\-event volatility\. This progressive decay strongly aligns with the financial theory of information half\-life: the qualitative and quantitative signals extracted directly from an earnings call are most potent in the immediate short term[Hafez et al\. \(2021\)](https://arxiv.org/html/2609.38523#bib.bib30);[Molinaro \(2026\)](https://arxiv.org/html/2609.38523#bib.bib31)\. As the prediction window expands to a full month, the initial catalyst of the earnings signal is rapidly diluted and overwhelmed by unpredictable market factors\. This confirms that forecasting long\-term variance is inherently unfeasible for current LLM architectures\.

To further validate the information half\-life theory, we map the Pearson correlation coefficients for the volatility tasks across the 2019–2022 timeline in Figure[4](https://arxiv.org/html/2609.38523#S6.F4)\. This visualization reinforces the theory by revealing a clear temporal degradation as the prediction horizon expands\. For instance, when observing the proprietary models across post\-event volatility windows, GPT and Claude achieve strong correlations for the 3\-day and 7\-day intervals \(fut\_3d and fut\_7d\); however, this predictive capability steadily deteriorates at the 15\-day and 30\-day marks \(fut\_15d and fut\_30d\)\. An isolated exception occurs with Claude 4\.8, which exhibits unusually high performance across all post\-event volatility tasks specifically in the year 2021\. Despite this outlier, the overarching trend of progressive decay strongly aligns with our findings in Figure[3](https://arxiv.org/html/2609.38523#S6.F3), reaffirming that the predictive signal contained within a corporate disclosure is rapidly diluted by unpredictable market factors over extended temporal windows\. More results analysis is in Appendix[C](https://arxiv.org/html/2609.38523#A3)\.

### 6\.6Temporal Degradation on Direction Accuracy

To evaluate how model performance changes over time, we analyze the annual Directional Accuracy trends from 2019 to 2022\. As illustrated in Figure[5](https://arxiv.org/html/2609.38523#S6.F5), there is a consistent decline in Directional Accuracy from 2019 to 2022 across the return\-based tasks \(ret\_eom, ret\_l5m, and ret\_ldm\)\. In 2019, several top\-performing models tightly cluster near peak accuracies between 70% and 90%\. However, this performance decreases and drops to a DA range of 30% to 40% in 2022\. This highlights a lack of temporal generalization in current LLMs\. It suggests that post\-pandemic market dynamics introduced increased market complexity that reduced the models’ predictive capabilities\.

## 7Conclusion

In this work, we introduced MM\-FinEval, a multimodal benchmark for evaluating LLMs across 12 financial tasks using 2,045 S&P 500 earnings conference calls spanning a 2019–2022 period\. The three modalities inputs in MM\-FinEval reflect the analysis process of expert human financial analysts, who must synthesize not only what executives disclose but how they present it to execute complex financial reasoning\. Our extensive baseline evaluation across Image\-Text, Audio\-Text, and Any\-to\-Any configurations exposes the boundary where pure language modeling must transition into multi\-sensory, calibrated quantitative reasoning\. We hope this dataset serves as a benchmark for the AI and finance communities, supporting future work toward multimodal financial analysis for economic forecasting\.

## References

- Bessembinderet al\.\(2008\)H\. Bessembinder, K\. M\. Kahle, W\. F\. Maxwell, and D\. XuMeasuring abnormal bond performance\.The Review of Financial Studies22\(10\),pp\. 4219–4258\.Cited by:[§3\.1](https://arxiv.org/html/2609.38523#S3.SS1.SSS0.Px3.p1.1)\.
- Chenet al\.\(2021\)Z\. Chen, W\. Chen, C\. Smiley, S\. Shah, I\. Borova, D\. Langdon, R\. Moussa, M\. Beane, T\. Huang, B\. R\. Routledge,et al\.Finqa: a dataset of numerical reasoning over financial data\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,pp\. 3697–3711\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Del Rioet al\.\(2021\)M\. Del Rio, N\. Delworth, R\. Westerman, M\. Huang, N\. Bhandari, J\. Palakapilly, Q\. McNamara, J\. Dong, P\. Zelasko, and M\. JettéEarnings\-21: a practical benchmark for asr in the wild\.arXiv preprint arXiv:2104\.11348\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p2.1)\.
- Donget al\.\(2024\)Z\. Dong, X\. Fan, and Z\. PengFnspid: a comprehensive financial news dataset in time series\.InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 4918–4927\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Duanet al\.\(2018\)J\. Duan, Y\. Zhang, X\. Ding, C\. Y\. Chang, and T\. LiuLearning target\-specific representations of financial news documents for cumulative abnormal return prediction\.InProceedings of the 27th international conference on computational linguistics,pp\. 2823–2833\.Cited by:[§3\.1](https://arxiv.org/html/2609.38523#S3.SS1.SSS0.Px1.p1.1)\.
- Hafezet al\.\(2021\)P\. Hafez, R\. Matas Navarro, F\. Gomez, and M\. KangrgaThree trading signals generated systematically from earnings conference calls\.Available at SSRN 4323558\.Cited by:[§6\.5](https://arxiv.org/html/2609.38523#S6.SS5.p3.1)\.
- He and Zhang \(2025\)Z\. He and C\. ZhangCross\-sectional spillovers of earnings surprises and asset price anomalies\.Available at SSRN 5415255\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Hirshleiferet al\.\(2009\)D\. Hirshleifer, S\. S\. Lim, and S\. H\. TeohDriven to distraction: extraneous events and underreaction to earnings news\.The Journal of Finance64\(5\),pp\. 2289–2325\.Cited by:[§3\.1](https://arxiv.org/html/2609.38523#S3.SS1.SSS0.Px1.p1.1)\.
- Huanget al\.\(2023\)A\. H\. Huang, H\. Wang, and Y\. YangFinBERT: a large language model for extracting information from financial text\.Contemporary Accounting Research40\(2\),pp\. 806–841\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p1.1)\.
- Huanget al\.\(2024\)R\. Huang, M\. Yarmohammadi, J\. Trmal, J\. Liu, D\. Raj, L\. P\. Garcia, A\. Ivanov, P\. Ehlen, M\. Yu, A\. Rastrow, D\. Povey, and S\. KhudanpurConEC: earnings call dataset with real\-world contexts for benchmarking contextual speech recognition\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 3700–3706\.External Links:[Link](https://aclanthology.org/2024.lrec-main.328/)Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p2.1)\.
- Jianget al\.\(2023\)A\. Q\. Jiang, A\. Sablayrolles, A\. Mensch, C\. Bamford, D\. S\. Chaplot, D\. de las Casas, F\. Bressand, G\. Lengyel, G\. Lample, L\. Saulnier, L\. R\. Lavaud, M\. Lachaux, P\. Stock, T\. L\. Scao, T\. Lavril, T\. Wang, T\. Lacroix, and W\. E\. SayedMistral 7b\.External Links:2310\.06825,[Link](https://arxiv.org/abs/2310.06825)Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Jostovaet al\.\(2013\)G\. Jostova, S\. Nikolova, A\. Philipov, and C\. W\. StahelMomentum in corporate bond returns\.The Review of Financial Studies26\(7\),pp\. 1649–1693\.Cited by:[§3\.1](https://arxiv.org/html/2609.38523#S3.SS1.SSS0.Px3.p1.1)\.
- Koganet al\.\(2009\)S\. Kogan, D\. Levin, B\. R\. Routledge, J\. S\. Sagi, and N\. A\. SmithPredicting risk from financial reports with regression\.InProceedings of human language technologies: the 2009 annual conference of the North American Chapter of the Association for Computational Linguistics,pp\. 272–280\.Cited by:[§3\.1](https://arxiv.org/html/2609.38523#S3.SS1.SSS0.Px2.p1.1)\.
- Konstantinidiset al\.\(2024\)T\. Konstantinidis, G\. Iacovides, M\. Xu, T\. G\. Constantinides, and D\. MandicFinllama: financial sentiment classification for algorithmic trading applications\.arXiv preprint arXiv:2403\.12285\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Laiet al\.\(2025\)V\. Lai, M\. Krumdick, C\. Lovering, V\. Reddy, C\. Schmidt, and C\. TannerSec\-qa: a systematic evaluation corpus for financial qa\.InProceedings of The 10th Workshop on Financial Technology and Natural Language Processing,pp\. 221–236\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Liet al\.\(2020\)J\. Li, L\. Yang, B\. Smyth, and R\. DongMaec: a multimodal aligned earnings conference call dataset for financial risk prediction\.InProceedings of the 29th ACM International Conference on Information & Knowledge Management,pp\. 3063–3070\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p2.1)\.
- Liet al\.\(2023\)Y\. Li, S\. Wang, H\. Ding, and H\. ChenLarge language models in finance: a survey\.InProceedings of the fourth ACM international conference on AI in finance,pp\. 374–382\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p1.1)\.
- Liang and Carrasco Kind \(2025\)Q\. Liang and M\. Carrasco KindHow does managers’ willingness to disclose affect analysts’ earning forecasts\-a measurement by llms\.Available at SSRN 5199752\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Liuet al\.\(2025a\)A\. H\. Liu, A\. Ehrenberg, A\. Lo, C\. Denoix, C\. Barreau, G\. Lample, J\. Delignon, K\. R\. Chandu, P\. von Platen, P\. R\. Muddireddy,et al\.Voxtral\.arXiv preprint arXiv:2507\.13264\.Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Liuet al\.\(2025b\)Z\. Liu, X\. Guo, Z\. Yang, F\. Lou, L\. Zeng, J\. Niu, M\. Li, Q\. Qi, Z\. Liu, Y\. Han,et al\.Fin\-r1: a large language model for financial reasoning through reinforcement learning\.arXiv preprint arXiv:2503\.16252\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Luet al\.\(2024\)Y\. Lu, Y\. Hu, H\. Foroosh, W\. Jin, and F\. LiuSTRUX: an llm for decision\-making with structured explanations\.arXiv preprint arXiv:2410\.12583\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Luoet al\.\(2025\)J\. Luo, Z\. Kou, L\. Yang, X\. Luo, J\. Huang, Z\. Xiao, J\. Peng, C\. Liu, J\. Ji, X\. Liu,et al\.Finmme: benchmark dataset for financial multi\-modal reasoning evaluation\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 29465–29489\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Magomereet al\.\(2025\)J\. Magomere, E\. Kochkina, S\. Mensah, S\. Kaur, and C\. SmileyFinNLI: novel dataset for multi\-genre financial natural language inference benchmarking\.InFindings of the Association for Computational Linguistics: NAACL 2025,L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),Albuquerque, New Mexico,pp\. 4545–4568\.External Links:[Link](https://aclanthology.org/2025.findings-naacl.257/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-naacl.257),ISBN 979\-8\-89176\-195\-7Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Mathuret al\.\(2022\)P\. Mathur, M\. Goyal, R\. Sawhney, R\. Mathur, J\. L\. Leidner, F\. Dernoncourt, and D\. ManochaDocfin: multimodal financial prediction and bias mitigation using semi\-structured documents\.InFindings of the Association for Computational Linguistics: EMNLP 2022,pp\. 1933–1940\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Molinaro \(2026\)M\. MolinaroDo earnings call transcripts predict post\-announcement returns?\.Available at SSRN 6695758\.Cited by:[§6\.5](https://arxiv.org/html/2609.38523#S6.SS5.p3.1)\.
- Mukherjeeet al\.\(2022\)R\. Mukherjee, A\. Bohra, A\. Banerjee, S\. Sharma, M\. Hegde, A\. Shaikh, S\. Shrivastava, K\. Dasgupta, N\. Ganguly, S\. Ghosh, and P\. GoyalECTSum: a new benchmark dataset for bullet point summarization of long earnings call transcripts\.InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing,Y\. Goldberg, Z\. Kozareva, and Y\. Zhang \(Eds\.\),Abu Dhabi, United Arab Emirates,pp\. 10893–10906\.External Links:[Link](https://aclanthology.org/2022.emnlp-main.748/),[Document](https://dx.doi.org/10.18653/v1/2022.emnlp-main.748)Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Pardawalaet al\.\(2024\)H\. Pardawala, S\. Sukhani, A\. Shah, V\. Kejriwal, A\. Pillai, R\. Bhasin, A\. DiBiasio, T\. Mandapati, D\. Adha, and S\. ChavaSubjECTive\-qa: measuring subjectivity in earnings call transcripts’ qa through six\-dimensional feature analysis\.Advances in Neural Information Processing Systems37,pp\. 59342–59372\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Qin and Yang \(2019\)Y\. Qin and Y\. YangWhat you say and how you say it matters: predicting stock volatility using verbal and vocal cues\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,pp\. 390–401\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p2.1)\.
- Qwen Team \(2026a\)Qwen TeamQwen3\.5: towards native multimodal agents\.External Links:[Link](https://qwen.ai/blog?id=qwen3.5)Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Qwen Team \(2026b\)Qwen TeamQwen3\.6\-35B\-A3B: agentic coding power, now open to all\.External Links:[Link](https://qwen.ai/blog?id=qwen3.6-35b-a3b)Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Sawhneyet al\.\(2020a\)R\. Sawhney, A\. Aggarwal, P\. Khanna, P\. Mathur, T\. Jain, and R\. R\. ShahRisk forecasting from earnings calls acoustics and network correlations\.\.InINTERSPEECH,pp\. 2307–2311\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Sawhneyet al\.\(2020b\)R\. Sawhney, P\. Khanna, A\. Aggarwal, T\. Jain, P\. Mathur, and R\. ShahVolTAGE: volatility forecasting via text audio fusion with graph convolution networks for earnings calls\.InProceedings of the 2020 conference on empirical methods in natural language processing \(EMNLP\),pp\. 8001–8013\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Sawhneyet al\.\(2020c\)R\. Sawhney, P\. Mathur, A\. Mangal, P\. Khanna, R\. R\. Shah, and R\. ZimmermannMultimodal multi\-task financial risk forecasting\.InProceedings of the 28th ACM international conference on multimedia,pp\. 456–465\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Shahet al\.\(2023\)A\. Shah, S\. Paturi, and S\. ChavaTrillion dollar words: a new financial dataset, task & market analysis\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 6664–6679\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Shuet al\.\(2025a\)D\. Shu, Y\. Liu, H\. Zhang, and M\. DuFinCall\-surprise: a large scale multi\-modal benchmark for earning surprise prediction\.arXiv preprint arXiv:2510\.03965\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p2.1)\.
- Shuet al\.\(2025b\)D\. Shu, H\. Yuan, Y\. Wang, Y\. Liu, H\. Zhang, H\. Zhao, and M\. DuFinchart\-bench: benchmarking financial chart comprehension in vision\-language models\.arXiv preprint arXiv:2507\.14823\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.
- Teamet al\.\(2024\)G\. Team, T\. Mesnard, C\. Hardin, R\. Dadashi, S\. Bhupatiraju, S\. Pathak, L\. Sifre, M\. Rivière, M\. S\. Kale, J\. Love,et al\.Gemma: open models based on gemini research and technology\.arXiv preprint arXiv:2403\.08295\.Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Team \(2026a\)G\. TeamGemma 4 technical report\.External Links:2607\.02770,[Link](https://arxiv.org/abs/2607.02770)Cited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Team \(2026b\)O\. TeamMOSS\-audio technical report\.Note:[https://github\.com/OpenMOSS/MOSS\-Audio](https://github.com/OpenMOSS/MOSS-Audio)GitHub repositoryCited by:[§5\.1](https://arxiv.org/html/2609.38523#S5.SS1.p2.1)\.
- Wuet al\.\(2023\)S\. Wu, O\. Irsoy, S\. Lu, V\. Dabravolski, M\. Dredze, S\. Gehrmann, P\. Kambadur, D\. Rosenberg, and G\. MannBloomberggpt: a large language model for finance\.arXiv preprint arXiv:2303\.17564\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p1.1)\.
- Yanget al\.\(2023\)H\. Yang, X\. Liu, and C\. D\. WangFingpt: open\-source financial large language models\.arXiv preprint arXiv:2306\.06031\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p1.1)\.
- Yuet al\.\(2025\)D\. Yu, Z\. Liu, and H\. HeSame company, same signal: the role of identity in earnings call transcripts\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 18403–18422\.Cited by:[§2\.1](https://arxiv.org/html/2609.38523#S2.SS1.p1.1)\.
- Zhuet al\.\(2025\)Y\. Zhu, X\. Liu, and O\. R\. L\. ShengPost\-earnings\-announcement drift prediction: leveraging postevent investor responses with multitask learning\.Information Systems Research36\(4\),pp\. 2191–2212\.Cited by:[§1](https://arxiv.org/html/2609.38523#S1.p2.1)\.

## Appendix AData Analysis

In Figure[2](https://arxiv.org/html/2609.38523#S4.F2), we detail the label statistics, including the mean, standard deviation, minimum, and maximum for all 12 financial task labels\. An analysis of these distributions explicitly confirms the structural differences between discrete return tasks and continuous volatility tasks\.

The return\-based tasks \(car, ret\_eom, ret\_l5m, and ret\_ldm\) exhibit a classic zero\-centered financial return distribution\. The means for these labels hover barely above zero, ranging from 0\.001620 to 0\.003443\. In contrast, their standard deviations are substantially larger, ranging from 0\.024981 to 0\.054573\. This high variance relative to the mean confirms a notoriously low signal\-to\-noise ratio inherent to predicting market direction\. Furthermore, the minimum and maximum values demonstrate that returns fluctuate wildly in both directions\. For instance, the Cumulative Abnormal Return \(car\) spans from a minimum of \-0\.261699 to a maximum of 0\.312476, illustrating the extreme, bidirectional stochastic shocks that models must account for\.

Conversely, the volatility\-based tasks display a fundamentally different statistical profile\. Most notably, the minimum values across all eight volatility labels are strictly positive \(e\.g\., the lowest absolute minimum is 0\.000288 for past\_3d\)\. This directly reflects the hard non\-negativity constraint of financial volatility\. The means are significantly higher and tightly clustered between 0\.017588 and 0\.020292\. Additionally, the standard deviations \(0\.009554 to 0\.015334\) are much smaller relative to their respective means when compared to the return labels\. This points to a tighter, more structured distribution characterized by volatility clustering, making it a more autoregressive and stable phenomenon\.

These distinct distributional characteristics mathematically explain the performance divergence observed in our benchmark results\. Return tasks are heavily obscured by symmetric, zero\-centered noise, making them highly resistant to accurate quantitative regression\. Volatility tasks, however, present a structured, bounded, and clustered target that is quantitatively easier for models to map, perfectly mirroring established financial econometrics\.

## Appendix BPrompt Used in Our Experiments

### B\.1CAR Task

Image\-Text ModelsYou are an expert quantitative equity analyst\. Your task is to predict the stock’s Cumulative Abnormal Return \(CAR\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the Cumulative Abnormal Return \(CAR\) associated with this earnings event\. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your CAR prediction must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ CAR decimal \]\]

Audio\-Text ModelsYou are an expert quantitative equity analyst\. Your task is to predict the stock’s Cumulative Abnormal Return \(CAR\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the Cumulative Abnormal Return \(CAR\) associated with this earnings event\. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your CAR prediction must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ CAR decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative equity analyst\. Your task is to predict the stock’s Cumulative Abnormal Return \(CAR\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the Cumulative Abnormal Return \(CAR\) associated with this earnings event\. CAR measures the sum of the differences between the stock’s actual return and its expected return over the event window\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your CAR prediction must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ CAR decimal \]\]

### B\.2ret\_eom

Image\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the End\-of\-Month corporate bond return \(ret\_eom\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_eom \(End\-of\-Month corporate bond return\)\. It represents the percentage gain or loss of an investment or business value over the course of one month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_eom prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_eom decimal \]\]

Audio\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the End\-of\-Month corporate bond return \(ret\_eom\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_eom \(End\-of\-Month corporate bond return\)\. It represents the percentage gain or loss of an investment or business value over the course of one month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_eom prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_eom decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the End\-of\-Month corporate bond return \(ret\_eom\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_eom \(End\-of\-Month corporate bond return\)\. It represents the percentage gain or loss of an investment or business value over the course of one month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_eom prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_eom decimal \]\]

### B\.3ret\_l5m

Image\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return for the last 5 days of the month \(ret\_l5m\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_l5m \(Return Last 5 Days of Month\)\. It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_l5m prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_l5m decimal \]\]

Audio\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return for the last 5 days of the month \(ret\_l5m\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_l5m \(Return Last 5 Days of Month\)\. It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_l5m prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_l5m decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return for the last 5 days of the month \(ret\_l5m\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_l5m \(Return Last 5 Days of Month\)\. It represents the percentage gain or loss of the corporate bond during the final 5 trading days of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_l5m prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_l5m decimal \]\]

### B\.4ret\_ldm

Image\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return specifically on the last trading day of the month \(ret\_ldm\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_ldm \(Return Last Day of Month\)\. It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_ldm prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_ldm decimal \]\]

Audio\-Text ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return specifically on the last trading day of the month \(ret\_ldm\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_ldm \(Return Last Day of Month\)\. It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_ldm prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_ldm decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative fixed\-income analyst\. Your task is to predict the corporate bond return specifically on the last trading day of the month \(ret\_ldm\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the ret\_ldm \(Return Last Day of Month\)\. It represents the percentage gain or loss of the corporate bond exactly on the final trading day of the month\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your ret\_ldm prediction must be represented as a decimal \(e\.g\., 0\.052 or \-0\.014\)\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ ret\_ldm decimal \]\]

### B\.5vol\_fut\_3d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 3\-day volatility \(vol\_fut\_3d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_3d \(Volatility Future 3 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_3d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 3\-day volatility \(vol\_fut\_3d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_3d \(Volatility Future 3 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_3d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 3\-day volatility \(vol\_fut\_3d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_3d \(Volatility Future 3 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 3 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_3d decimal \]\]

### B\.6vol\_fut\_7d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 7\-day volatility \(vol\_fut\_7d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_7d \(Volatility Future 7 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_7d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 7\-day volatility \(vol\_fut\_7d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_7d \(Volatility Future 7 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_7d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 7\-day volatility \(vol\_fut\_7d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_7d \(Volatility Future 7 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 7 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_7d decimal \]\]

### B\.7vol\_fut\_15d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 15\-day volatility \(vol\_fut\_15d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_15d \(Volatility Future 15 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_15d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 15\-day volatility \(vol\_fut\_15d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_15d \(Volatility Future 15 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_15d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 15\-day volatility \(vol\_fut\_15d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_15d \(Volatility Future 15 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 15 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_15d decimal \]\]

### B\.8vol\_fut\_30d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 30\-day volatility \(vol\_fut\_30d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_30d \(Volatility Future 30 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_30d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 30\-day volatility \(vol\_fut\_30d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_30d \(Volatility Future 30 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_30d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to predict the post\-event 30\-day volatility \(vol\_fut\_30d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Predict the vol\_fut\_30d \(Volatility Future 30 Days\)\. This represents the expected price turbulence, uncertainty, and fluctuation of the asset over the 30 trading days immediately following this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_fut\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_fut\_30d decimal \]\]

### B\.9vol\_past\_3d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 3\-day volatility \(vol\_past\_3d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_3d \(Volatility Past 3 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 3 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_3d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 3\-day volatility \(vol\_past\_3d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_3d \(Volatility Past 3 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 3 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_3d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 3\-day volatility \(vol\_past\_3d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_3d \(Volatility Past 3 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 3 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_3d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_3d decimal \]\]

### B\.10vol\_past\_7d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 7\-day volatility \(vol\_past\_7d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_7d \(Volatility Past 7 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 7 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_7d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 7\-day volatility \(vol\_past\_7d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_7d \(Volatility Past 7 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 7 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_7d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 7\-day volatility \(vol\_past\_7d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_7d \(Volatility Past 7 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 7 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_7d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_7d decimal \]\]

### B\.11vol\_past\_15d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 15\-day volatility \(vol\_past\_15d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_15d \(Volatility Past 15 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 15 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_15d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 15\-day volatility \(vol\_past\_15d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_15d \(Volatility Past 15 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 15 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_15d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 15\-day volatility \(vol\_past\_15d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_15d \(Volatility Past 15 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 15 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_15d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_15d decimal \]\]

### B\.12vol\_past\_30d

Image\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 30\-day volatility \(vol\_past\_30d\) for a company based on its earnings call transcript and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_30d \(Volatility Past 30 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 30 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_30d decimal \]\]

Audio\-Text ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 30\-day volatility \(vol\_past\_30d\) for a company based on its earnings call transcript and earnings call mp3 audio\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_30d \(Volatility Past 30 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 30 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_30d decimal \]\]

Any\-to\-Any ModelsYou are an expert quantitative analyst\. Your task is to estimate the pre\-event 30\-day volatility \(vol\_past\_30d\) for a company based on its earnings call transcript, earnings call audio, and the accompanying presentation slides\.Transcript:\{text\_transcript\}Instructions:1\. Estimate the vol\_past\_30d \(Volatility Past 30 Days\)\. This represents the realized volatility \(price turbulence and uncertainty\) of the asset over the 30 trading days immediately preceding this earnings call\.2\. Provide 1\-2 sentences explaining the key driver\.3\. Your vol\_past\_30d estimate must be represented as a decimal\.\*\*Answer format\*\*Rationale = \[\[ 1\-2 sentences rationale \]\]Prediction = \[\[ vol\_past\_30d decimal \]\]

![Refer to caption](https://arxiv.org/html/2609.38523v1/pearson_trend.png)Figure 6:Annual Pearson Correlation Trends \(2019–2022\) on the MM\-FinEval Benchmark\. The heatmap illustrates how models’ ability to capture linear relationships shifts over time\.

## Appendix CPearson Correlation Temporal Observations

We evaluated the Pearson correlation coefficient across the 2019–2022 timeline\. Figure[6](https://arxiv.org/html/2609.38523#A2.F6)indicates that proprietary models improve their ability to capture linear relationships over time\. For example, on the ret\_eom task, GPT\-5\.5 exhibits a strict upward trajectory, scaling from a correlation of 0\.309 in 2019 to 0\.581 by 2022\. Similarly, Claude 4\.8 improves from 0\.144 in 2019 to 0\.437 in 2022 on the exact same task\. This temporal strengthening suggests that massive, closed\-source models are highly adept at mapping evolving macroeconomic environments to relative asset rankings\. In contrast, open\-source models demonstrate no such temporal adaptation\. Models like Qwen 3\.5 and Mistral 3 remain near\-zero correlations on ret\_eom across the four\-year span\.

Table 3:Annual MAE Trends \(2019–2022\) on the MM\-FinEval Benchmark\.
## Appendix DExamples of Model Responses

Qwen3\.5 2BRationale = The Cumulative Abnormal Return \(CAR\) is predicted to be positive, driven by the company’s achievement of all\-time record quarterly and full\-year operating income and operating ratio, which significantly outperformed analyst expectations and the broader market\.Prediction = \[\[ 0\.045 \]\]

Qwen3\.5 4BRationale = \[\[ The company reported a 22% decline in GAAP earnings per share due to significant one\-time costs, including severance charges and litigation expenses, which negatively impacted the market’s sentiment despite underlying non\-GAAP growth\. The management’s focus on cost\-cutting and the uncertainty surrounding future market conditions likely led to a negative market reaction\.\]\]Prediction = \[\[ \-0\.035 \]\]

Qwen3\.5 9BRationale = The company delivered solid full\-year results with revenue growth, margin expansion, and strong free cash flow, but the stock price likely already priced in the positive momentum from the second half of the year\. Additionally, the guidance for 2019 EPS of $13\.90 was presented as a f̈loorr̈ather than a target, and the company explicitly noted that the Red Hat acquisition \(a major value driver\) is excluded from 2019 financial expectations due to timing uncertainty, which may limit the upside surprise in the immediate reaction\.Prediction = \[\[ 0\.012 \]\]

Qwen3\.5 27BRationale = \[\[ The company reported a 22% year\-over\-year decline in GAAP earnings per share to $0\.84, missing expectations due to significant one\-time charges for severance, litigation, and headquarters relocation totaling $0\.16 per share\. Although non\-GAAP earnings grew 9%, the market reaction is likely to be negative due to the headline GAAP miss and the explicit guidance that equity markets will not be a significant contributor to revenue growth in 2019\. \]\]Prediction = \[\[ \-0\.035 \]\]

Qwen3\.5 35BRationale = PNC delivered a strong 2020 with record revenue and positive operating leverage, but the market reaction is likely tempered by the guidance for a modest decline in Net Interest Income \(NII\) for 2021 and the significant drag on reserves \($3\.2 billion provision\) due to the pandemic, resulting in a neutral to slightly negative market response\.Prediction = \-0\.012

Qwen3\.6 27BRationale = \[\[ The company significantly raised its full\-year 2021 revenue guidance to a range of $5\.5\-$6\.0 billion and projected a substantial expansion in operating margins to 10\.4%\-11\.0%, driven by strong net order growth and a robust backlog\. Additionally, the firm highlighted a strong balance sheet with a low debt\-to\-capital ratio and a clear path to double\-digit community count growth, signaling strong future profitability\. \]\]Prediction = \[\[ 0\.065 \]\]

Qwen3\.6 35BRationale = BlackRock reported record full\-year revenue growth of 20% and operating income growth of 19%, significantly outperforming expectations with record net inflows of $540 billion and a 10% increase in AUM to $10 trillion\. The company also announced a strong capital return program, including an 18% dividend increase and $1\.5 billion in share repurchases, signaling high confidence in future cash flows and growth\.Prediction = \[\[ 0\.035 \]\]

Gemma3nRationale = \[\[ The earnings call presented a strong Q3 performance with solid growth across key products like Opdivo and Eliquis, and positive momentum from new launches like Zeposia and Abecma\. The company reaffirmed its financial guidance and expressed confidence in its ability to navigate the challenges of generic erosion while continuing to invest in innovation, which is likely to be viewed favorably by investors\. \]\]Prediction = \[\[ 0\.06 \]\]

Mistral3 3BRationale = \[The strong execution in 2018, with significant improvements in ROTCE \(14\.1%\), efficiency ratio \(56\.7%\), and underlying EPS growth, alongside disciplined balance sheet growth and robust credit quality, suggests strong investor confidence\. The company’s strategic initiatives like TOP programs, balance sheet optimization, and digital transformation, combined with a positive economic outlook and prudent risk management, should drive positive abnormal returns\. The stock’s current valuation discount relative to peers, especially in a recession\-sensitive sector, and the company’s ability to sustain growth through organic and strategic acquisitions, are likely to support a positive CAR\.\]Prediction = \[0\.065\]

Mistral3 8BRationale =Capital One’s fourth quarter 2019 earnings call and presentation highlighted several positive yet nuanced financial developments\. The company reported strong loan growth \(7% year\-over\-year\), solid revenue growth \(7%\), and a stable credit environment, though offset by increased provisions for credit losses \(up 11%\) and the transition to CECL, which added $2\.9B to the allowance\. The Walmart partnership launch, digital transformation progress, and moderate marketing expense growth suggest sustained growth momentum, but the CECL impact and higher provisioning could dampen near\-term profitability expectations\. The net charge\-off rate remained stable at 1\.93%, and the company’s disciplined capital management and strategic investments in technology and marketing indicate long\-term value creation\.Prediction = \[\[ 0\.025 \]\]",

Mistral3 14BRationale = \[\[BNY Mellon’s earnings call highlighted strong organic growth, innovation in digital assets and real\-time payments, and a robust capital position, which should positively surprise investors\. The guidance for double\-digit EPS growth in 2022, driven by higher net interest revenue and fee revenue recovery, further supports a positive market reaction\.\]\]Prediction = \[\[0\.035\]\]

GPT 5\.4Rationale = \[\[ Kimberly\-Clark delivered a clear 2020 beat with strong EPS, margin expansion, cash flow, and improved market share, but the market reaction is likely tempered by a cautious 2021 outlook featuring only 1\-2% organic growth and significant commodity inflation that is expected to pressure margins\. The net effect should be modestly positive because execution and restructuring/cost savings were strong, yet guidance implies a tougher earnings quality and margin backdrop ahead\. \]\]Prediction = \[\[ 0\.012 \]\]

GPT 5\.5Rationale = \[\[ Despite record full\-year results, the event was dominated by investor concern over a Q4 EPS miss driven by sharply higher compensation/non\-comp expenses, wage inflation, and weaker sequential trading performance\. Management’s commentary on continued investment spending, capital headwinds, and limited near\-term buybacks likely reinforced negative revisions to forward profitability\. \]\]Prediction = \[\[ \-0\.065 \]\]

Claude 4\.6Rationale = \[\[ PMI delivered a strong beat with Q2 adjusted diluted EPS of $1\.57 \(\+17\.8% organic\), raised full\-year EPS guidance to $5\.97\-$6\.07 \(implying \+15\-17% in dollar terms\), confirmed a $5\-7B share buyback program, and announced the imminent IQOS ILUMA Japan launch all meaningfully positive catalysts that likely exceeded consensus expectations and drove a positive abnormal return\. \]\]Prediction = \[\[ 0\.038 \]\]

Claude 4\.8Rationale = \[\[ WBA delivered a strong Q1 beat \(adjusted EPS $1\.68, up 53%, well ahead of expectations\) driven by COVID vaccinations/testing and robust retail comps, plus raised full\-year guidance, which should be received positively despite headwinds from shrink, labor investments, and softer underlying scripts\. \]\]Prediction = \[\[ 0\.045 \]\]

Voxtral 24B\*\*Answer\*\*Rationale = The company’s strong financial performance, with a 32% increase in comparable basis EPS excluding canopy equity earnings, and the announcement of a $2 billion share repurchase program, are the key drivers for the stock’s CAR\. The company’s commitment to returning $5 billion in value to shareholders through fiscal 23 also supports a positive outlook\.Prediction = 0\.035

MossRationale = \[\[ This earnings call seems to be a ’surprising positive’ event which can lead to a strong positive CAR\. It may provide a signal that the outlook for the company is improving, leading to an increase in the company’s stock price\. \]\]Prediction = 0\.123

Gemma4 E2BRationale = \[\[ The company reported strong retail sales growth in Q3, driven by improved mix and digital sales, and provided guidance indicating a strong Q3 and a cautious outlook for Q4\. The financial results show a positive trend in retail sales and margin expansion, which are positive indicators for the stock\. \]\]Prediction = \[\[ 0\.025 \]\]

Gemma4 E4BRationale = \[\[ The earnings call presents a positive outlook with strong underlying sales growth and a significant increase in adjusted segment EBIT, which should lead to a positive market reaction\. The company also highlights strong free cash flow and a positive ESG momentum, which further supports a bullish sentiment\. \]\]Prediction = \[\[ 0\.035 \]\]