Source-linked AI summary
Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting
Jihoon Kwon, Lawrence Liu, Daekyung Park, Sumin Kim, Haverty Jack, Hoyoung Lee, Katherine Bjorkman, Josh McKenney, Peter Laurelli, Nicole Kagan, Zach Golkhou, Thorsten Neumann, Edward Tong, Pete Petersen, Yoon Kim, Alejandro Lopez-Lira, Yongjae Lee, Chanyeol Choi
TL;DR
The paper addresses how to use limited, heterogeneous alternative data in firm-level financial forecasting. It develops a two-agent in-context framework that screens informative firm-channel pairs and predicts revenue from financial history, alternative data, and earnings calls. Across four commercial channels, the resulting forecasts improve on forecasts without alternative data, feature-based baselines, and analyst consensus for most revenue-surprise events.
Problem
Alternative data have limited historical coverage and heterogeneous, firm- or channel-specific relevance, making them difficult to incorporate flexibly into conventional forecasting approaches.
Method
A two-agent in-context framework screens informative firm-channel pairs and then jointly interprets financial history, alternative data, and prior earnings-call transcripts to forecast revenue.
Results
Across four commercial alternative data channels, forecasts improve on those without alternative data, feature-based supervised baselines, and analyst consensus for a majority of revenue-surprise events.
Takeaways & Limitations
In-context reasoning provides a practical way to incorporate new structured and textual information sources into financial forecasting.
Takeaways & Limitations
The evaluation uses one tabular source and one textual source for each source type rather than multiple channels of the same type.
Abstract
from arXiv · showhide
When forecasting a firm's future financial performance, alternative data - data collected from non-traditional sources such as consumer transactions, web traffic, and prediction markets - can provide timely signals about firms' operating activities and broader market conditions. These signals may reveal information that is not captured by traditional public sources and can therefore provide complementary information for forecasting firms' future financial performance. However, firm-level alternative data often have limited historical coverage, are relevant only to specific prediction targets or subsets of firms, and are distributed across numerous heterogeneous channels, making them difficult to incorporate flexibly into conventional forecasting approaches. Meanwhile, large language models (LLMs) can interpret instructions, learn from in-context examples, and generate predictions by combining heterogeneous information without task-specific parameter updates. Motivated by this potential flexibility, we investigate whether an LLM can forecast firm performance by integrating alternative data with other financial information through in-context learning. We propose a two-agent framework that first identifies the firms for which each alternative data channel is likely to be informative and then predicts revenue using firm- and channel-specific context. We evaluate the framework across four commercial alternative data channels. In our experiments, adding alternative data in context alongside other financial information improves the LLM's forecasting relative to either source alone, and these forecasts are more accurate than those of standard forecasting baselines. These findings suggest that LLMs provide a flexible and practical approach to integrating alternative data with heterogeneous financial information.
1 Introduction
The paper asks whether LLMs can flexibly integrate heterogeneous alternative data into firm-level financial forecasting despite limited coverage and channel-specific relevance. It proposes in-context learning with a two-agent screening-and-prediction framework and finds consistent gains from combining alternative data with conventional financial information.
- Motivation: Alternative data offer timely signals about firms’ operating activity but are difficult to incorporate into conventional supervised forecasting models because of limited histories and heterogeneous channels.These constraints limit the number of firm-quarter observations and make task-specific model design difficult.
- Motivation: In-context learning lets one LLM infer forecasting tasks from instructions and examples without separate parameter updates for each channel or task.This flexibility motivates using LLMs to combine alternative data with financial information.
- Framework: The framework screens firm-channel pairs for likely relevance before jointly interpreting financial history, alternative data, and earnings-call transcripts to forecast revenue.The screening stage filters channels unlikely to reflect revenue-generating activity.
- Evidence: Across four commercial channels, adding alternative data alongside conventional financial information consistently improves forecasting relative to conventional information alone.The evaluation holds the underlying LLM and sample fixed across information conditions.
- Evidence: The framework improves on trained feature-based baselines and beats analyst consensus in a majority of revenue-surprise evaluation events.These comparisons support the practical forecasting value of the proposed approach.
2 Related Work
Prior work finds predictive signals in corporate disclosures, managerial language, and operational alternative data, while LLM research shows that in-context learning can support regression-like prediction. The paper builds on these strands by combining complementary information sources for financial forecasting.
- Financial forecasting from heterogeneous information sources: Alternative data predict revenue and forecast errors when they directly capture firms’ revenue-generating activities.Examples include POS transactions and digital traffic for firms whose websites support consumer transactions.
- Financial forecasting from heterogeneous information sources: Earnings calls reveal management’s expectations, whereas operational alternative data capture subsequently realized customer and business activity.The two sources describe the same business from different perspectives and at different times.
- In-context learning for regression and financial forecasting: In-context learning enables pretrained LLMs to infer tasks and perform regression from information supplied in the prompt without parameter updates.Prior controlled studies report regression-like learning from input-output examples.
3 Problem Setup
The paper formulates firm-quarter forecasting with baseline, structured alternative-data, and textual information sources, while always conditioning on information available before the outcome. It evaluates source combinations using normalized prediction loss and mean absolute error.
- 3.1 Data and Models: The forecasting unit is a firm-quarter with baseline history H, structured alternative data X, and textual earnings-call information Z.The formulation permits optional source sets while retaining baseline information in every condition.
- 3.1 Data and Models: The model compares H, HX, HZ, and HXZ conditions, where H is always present and X and Z are optional sources.This design isolates the contribution of each information source and their combination.
- 3.2 Task and Metrics: Only inputs known before outcome reporting are used to predict the firm’s performance outcome Y.This defines the information timing constraint for the forecasting task.
- 3.2 Task and Metrics: FVU is the primary normalized metric for comparing forecasts across targets, with lower values indicating better performance.An FVU of 0 denotes perfect prediction and 1 denotes predicting the target mean.
- 3.2 Task and Metrics: MAE supplements FVU by measuring typical absolute prediction error with less sensitivity to outliers than squared loss.FVU instead emphasizes large prediction errors through squared loss.
4 Method
The framework uses a screening agent to identify informative firm–channel pairs and a prediction agent to forecast revenue from financial, alternative-data, and transcript context.
- Motivation for screening: The two-agent framework screens firm–channel relevance before using selected alternative data for revenue forecasting.The screening decision is based on observable business characteristics and channel definitions, without future outcomes.
- Screening procedure: The screening agent labels each firm–channel pair INCLUDED or EXCLUDED using the firm’s name, business description, and channel description.When information is insufficient, it may retrieve company profiles or financial statements before deciding.
- Motivation for Tool-Augmented Prediction: Search tools provide firm- and channel-specific context because numerical channel values alone may not reveal their predictive meaning.Interpretation depends on how the channel relates to the firm’s revenue-generating activities.
- Prediction Agent: Table 2 compares Revenue-YoY forecasts using financial statements alone, financial statements with alternative data, financial statements with transcripts, and all three sources.Across four channels, the all-source condition generally outperforms the alternatives.
- Prediction Procedure: The prediction agent jointly reads financial history, alternative-data time series, and earnings-call transcripts to output the target revenue with a rationale.It consults search tools when background is needed to interpret revenue drivers or the channel’s relationship to them.
5 Experiments
The experiments evaluate a two-agent LLM forecasting framework across heterogeneous alternative-data channels, comparing information combinations, screening, tools, and supervised baselines. Across these tests, combining alternative data with financial history and transcripts improves forecasts, while screening and firm- or channel-specific context further support prediction.
- 5.2 Results: Across four alternative-data channels, combining financial history, alternative data, and transcripts generally yields the lowest prediction error, outperforming either source alone.For web traffic, FVU falls to 0.021 from 0.044 with H + Z and 0.124 with H + X.
- 5.2 Results: Adding alternative data often moves individual forecasts closer to realized revenue than text-only predictions, indicating integration of multiple information sources.Figure 3 compares estimates with and without alternative data and highlights the alternative-data arguments in the generated rationale.
- 5.2 Results: The full method outperforms historical averages, OLS, and gradient-boosted trees trained on the same information under limited data.In card spending, the method reports FVU 0.122 and MAE 2.49, versus OLS FVU 0.428 and GBT FVU 0.453.
- 5.2 Results: Against analyst consensus, the method achieves lower prediction error across every channel and wins in more than half of evaluation cases.For early consensus on web traffic, the method is closer to realized revenue in 81% of cases while also reducing RMSE.
- 5.2 Results: Screening strengthens the firm-specific relationship between alternative-data signals and revenue surprise by retaining firms whose revenues the channel directly captures.The correlation increases from 0.238 to 0.281 after screening.
- 5.2 Results: Firm- and channel-specific tool context improves prediction by helping the model interpret proprietary alternative-data signals.For web traffic, FVU decreases from 0.079 to 0.021 and MAE from 2.11 to 1.39 with search-tool access.
6 Conclusion
The paper studies whether general-purpose LLMs can incorporate firm-level alternative data into financial forecasting without task-specific design or supervised training. Its two-agent framework improves forecasts across four commercial channels and supports flexible integration of heterogeneous financial information.
- Across four commercial alternative data channels, the two-agent framework improves forecasts over versions without alternative data, supervised feature-based baselines, and analyst consensus for most revenue-surprise events.The framework first screens firm-channel relevance and then forecasts revenue using firm-specific financial, alternative, and textual information.
- The framework enables firm-level forecasting despite each alternative data channel providing only a limited number of observations.
- In-context reasoning provides a practical way to incorporate new structured and textual information sources into financial forecasting.
A.1 Additional Experiments
An additional experiment tests whether the main findings depend on GPT-5.5 by repeating the forecasting procedure with GLM-5.2 under a smaller, YoY-growth-only evaluation. The same performance ordering persists across all three evaluated channels.
- Experimental setup: The alternative-model experiment is designed to distinguish gains from the information-integration method from capabilities specific to GPT-5.5.
- Experimental setup: The replication preserves the main forecasting protocol while changing the model, knowledge-cutoff boundary, eligible sample sizes, and evaluated target.The eligible test samples fall to 131 card-spending, 27 web-traffic, and 58 foot-traffic firm-quarters, and evaluation is limited to revenue YoY growth.
- Results: With GLM-5.2, the full method achieves the lowest FVU and MAE in all three channels among the four source conditions.It also beats the historical average, OLS, and gradient-boosted trees by a wide margin.
A.2 Prompt Templates
The appendix provides the prompt templates used by the screening and prediction agents. Figures 6–8 show the system and user messages plus the structured outputs required for screening, YoY-growth prediction, and revenue-surprise prediction.
- Figures 6–8 document the screening prompt and prediction prompts for year-over-year growth and revenue-surprise targets.
- Each prompt figure shows the system message, user message, and structured output fields, with braced names serving as runtime placeholders.
- The screening prompt asks whether a specified alternative data channel is a dominant driver of a company’s total revenue.