Can Language Models Learn to Forecast Stock Prices

Jiacheng Guo; Suozhi Huang; Shuzhen Li; Yunlong Gao; Zerui Cheng; Jason Ge; Shushu Liang; Zihao Li

cs.CLSep 29, 2026

Can Language Models Learn to Forecast Stock Prices

Authors: Jiacheng Guo, Suozhi Huang, Shuzhen Li, Yunlong Gao, Zerui Cheng, Jason Ge, Shushu Liang, Zihao Li, +6 more

Organizations: Princeton University · InclusionAI · Harvard University · Stanford University

Abstract

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Figures & tables

Figure 1: Post-training brings a 4B model to performance comparable to frontier models on BETA. (a) Scores on the 240 scored test tasks improve from 20.94 to 37.94 to 43.31 through Base, SFT, and PPO. (b) Median calls across all 398 test tasks, including submission, change from 5 to 18 to 16. (c) AURA-4B ranks third among fifteen systems; the intermediate SFT checkpoint is also shown.

Evidence	Available observations	Analysis supported
Individual stock	OHLCV candles, quotes, technical indicators	Price trends, volume, volatility, and multiple resolutions
Relative performance	Rankings, relative returns, beta, correlation, residual returns	Position among peers and relative strength
Market context	Market/sector candles, breadth, symbol metadata	Sector conditions and the broader market

Table 1: Market information available in the sandbox. Full tool descriptions and task prompts appear in appendix B .

Property	Training	Test
Tasks	3,807	398
Distinct stocks	489	224
Decision dates	Feb. 2011–Dec. 2023	Feb. 2024–Mar. 2026
Scored evaluation tasks	—	240

Table 2: Benchmark split within a 499-equity market environment. The latest training label is realized on December 29, 2023; the earliest test decision is February 1, 2024.

Starting model
System	Score $S$	Direction $D$	Magnitude $M$	Later score
Qwen3-4B base	20.94	62.9	33.3	17.41
Frontier language models
glm-5.3	49.84	79.2	63.0	23.43
claude-fable-5	44.48	85.4	52.1	20.02
kimi-k3	36.28	80.4	45.1	21.84

Table 3: Existing models exhibit different strengths in direction and magnitude. Results for fourteen systems before adding AURA-4B.

S

is the joint forecast score,

D

is directional accuracy, and

M

is magnitude agreement conditional on a correct sign. Overall results use 240 tasks; the later-period column uses 71.

Checkpoint	Score	Relative gain vs. previous	Relative gain vs. Base
Base Qwen3-4B	20.94	—	—
SFT	37.94	+81.2%	+81.2%
AURA-4B (SFT + PPO)	43.31	+14.2%	+106.8%

Table 4: Training-stage results on the 240 scored test tasks. Gains are relative improvements.

Figure 2: AURA approaches GLM on easy and medium tasks and exceeds momentum on medium and hard tasks. (a) Scores for the three systems with recorded difficulty aggregates. (b) AURA scores higher than GLM on short horizons (1–10 days), while GLM scores higher on longer horizons (21–126 days). Sample sizes are shown below each group.

Figure 3: Post-training nearly doubles conditional magnitude agreement. The joint score factors as

S=D\,M/100

. Direction changes from 62.9 to 65.4, while conditional magnitude changes from 33.3 to 66.2.

Recorded behavior	Base	SFT	AURA-4B
Median tool calls	5	18	16
Median distinct retrieval tools	4	7	6
Ranking and market/sector query share	16.2%	31.4%	38.0%
Median absolute forecast	0.0120	0.0312	0.0420
Positive forecasts on scored tasks	80.4%	47.9%	58.8%

Table 5: Behavior across Base, SFT, and AURA-4B (SFT + PPO) on the same test tasks. The first four rows use all 398 tasks; positive-forecast shares use the 240 scored tasks. Call counts include the terminal submission; distinct retrieval tools and query shares exclude it. Query shares pool calls across tasks and count get_ranking and get_market_kline .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Action	Numerical information returned
get_kline	OHLCV candles at resolutions from one minute to one month.
get_quote	Session prices, daily VWAP, gaps, returns, volume ratio, and 52-week range.
get_market_kline	Market or sector candles and market-breadth statistics.
get_indicator	A caller-parameterized indicator time series.
list_indicators	Indicator names and default parameters.
get_ranking	Cross-sectional metric rankings, including the target stock’s rank.

Appendix

Table 6: The ten available actions: nine retrieval tools and one terminal prediction action.

System	Earlier ( $n=169$ )	Later ( $n=71$ )	Change (%)
AURA-4B	42.17	46.04	+9.2
Momentum extrapolation	41.99	42.01	+0.0
Momentum-20	29.77	36.37	+22.2
GBM classifier	32.32	33.13	+2.5
gpt-5.6-sol	37.06	25.08	-32.3
glm-5.3	60.94	23.43	-61.6

Appendix

Table 7: Calendar-period results for every system. Earlier: February 2024–July 2025. Later: August 2025–March 2026. Relative changes are computed from the rounded endpoints.

Figure 4: Absolute forecast errors for the recorded DOW task. The realized log return is

-0.0796

; errors are multiplied by 100.

Explore similar work

Reasoning through Verifiable Forecast Actions: Consistency-Grounded RL for Financial LLMs
May 21, 2026Jialin Chen, Aosong Feng, Harshit Verma +7Data-Driven Forecasting LLM Reasoning Strategies
Yale University · Arizona State University · University of Texas Rio Grande Valley
Financial Numerical Prediction and Allocation as Token Generation
Aug 10, 2026Xu Ouyang, Moontae LeePortfolio Construction
University of Illinois Chicago
Do Your Own Research: Learning to Forecast by Learning to Search
Oct 1, 2026Yusuf Afifi, Artur Kiulian, Anton Polishko +3Large Language Model Forecasting Data-Driven Forecasting
Future Principle

May 21, 2026, cs.LGJ/K move · Enter open · S save

Reasoning through Verifiable Forecast Actions: Consistency-Grounded RL for Financial LLMs

Jialin Chen, Aosong Feng, Harshit Verma, Siyi Gu+6

Yale University · Arizona State University · University of Texas Rio Grande Valley

Data-Driven Forecasting LLM Reasoning Strategies Financial Question Answering

Open paper AskPDF arXiv

Financial markets are characterized by extreme non-stationarity, low signal-to-noise ratios, and strong dependence on external information such as news, company fundamentals, and macroeconomic signals. Yet, existing approaches either abstract time-series into text or decouple forecasting from language-based reasoning, leading to a fundamental mismatch between qualitative reasoning and quantitative outcomes. To address this, we introduce StockR1, a time-series-enhanced LLM that unifies stock forecasting and financial reasoning through a verifiable forecast action. Based on a tool-call design, the model first emits a forecast action, which is a structured and interpretable representation of its qualitative market outlook. It then invokes a time-series decoder conditioned on this action to generate distributional future trajectories, leading to more informed question answering and financial reasoning. We optimize the full pipeline with reinforcement learning, where rewards jointly reflect answer validity, forecast accuracy, and consistency between generated actions and observed time-series dynamics. In addition, rewards are reweighted by a sample-level uncertainty scalar, encouraging the model to accommodate varying uncertainty in market dynamics. We evaluate StockR1 on financial question answering and stock forecasting over a large-scale 10-year benchmark. Our method consistently outperforms time-series baselines and general-purpose LLMs, improving reasoning accuracy by 17.7% (4B) and 25.9% (8B). These findings demonstrate that structuring the forecast actions establishes a powerful synergy between language reasoning and temporal prediction, enabling LLMs to reason through verifiable, interpretable, and numerically grounded decisions.

Abstract

Explore similar work

Reasoning through Verifiable Forecast Actions: Consistency-Grounded RL for Financial LLMs

Financial Numerical Prediction and Allocation as Token Generation

Do Your Own Research: Learning to Forecast by Learning to Search