PaperlayerView exact revision on arXiv

Source-linked AI summary

Can Language Models Learn to Forecast Stock Prices

Jiacheng Guo, Suozhi Huang, Shuzhen Li, Yunlong Gao, Zerui Cheng, Jason Ge, Shushu Liang, Zihao Li, Hao Lu, Ming Yin, Shilong Liu, Jiashuo Liu, Xu Kuang, Mengdi Wang

arXiv:2609.36914v1cs.CL

TL;DR

Post-training has improved language models on tasks with verifiable outcomes, but it is unclear whether it can improve stock forecasting when returns are noisy and relevant evidence is not obvious. The paper trains Qwen3-4B in a chronological tool-use forecasting benchmark with SFT and PPO. Its forecast score more than doubles, reaching performance comparable to frontier language models on the benchmark.

  • Problem

    Whether post-training can improve stock-price forecasting remains unclear because returns are noisy and the relevant information set is not obvious in advance.

    Source: §1 · p. 1

  • Method

    The authors train Qwen3-4B with SFT on tool-use demonstrations and then PPO using realized-return forecast scores in a chronological stock-price sandbox.

    Source: §1 · p. 2

  • Results

    20.94 to 43.31: post-training raises the forecast score, bringing AURA-4B to performance comparable to frontier language models on Beta.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Outcome-based post-training can substantially improve financial forecasting performance even when returns are noisy and the information-gathering strategy is unspecified in advance.

    Source: §1 · p. 3

    The reported forecast-score improvement is scoped to the outcome-selected historical tasks in this benchmark.

    Source: §7 · p. 11

Abstract

from arXiv · show

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.

1. Introduction

The paper asks whether post-training can improve stock forecasts when returns are noisy and the useful information set is unclear. On a chronological tool-use benchmark, it reports forecast-score gains from Base through SFT to PPO, reaching performance comparable to frontier models.

1. Introduction

  • Post-training gains on verifiable tasks motivate asking whether models can learn from historical outcomes to forecast the future, despite noisy returns and an unclear relevant information set.Forecast errors may reflect evidence choice, interpretation, or subsequent market variation, and the final score does not identify which observations were useful.Source: §1 · p. 1
  • 20.94→37.94→43.31: the forecast score rises from Base through SFT to PPO, bringing AURA-4B to performance comparable to frontier models on Beta.Figure 1 reports scores on 240 scored test tasks; both SFT and PPO contribute to the improvement.Source: Figure 1 · p. 2, §1 · pp. 2–3
  • Beta gives models a common tool-use interface to query price, volume, relative performance, and market context before submitting a numerical return forecast.It covers 499 US equities and horizons from one trading day to approximately six months; the test set has 398 tasks, 240 of which meet the scoring minimum-move criterion.Source: §1 · p. 2
  • Qwen3-4B is trained first with SFT on tool-use demonstrations, then with PPO rewards computed from realized returns, using a strict chronological train-test split.The Base, SFT, and SFT+PPO checkpoints separate demonstration-learning gains from additional outcome-optimization gains.Source: §1 · p. 2
  • Post-training changes investigation behavior and improves magnitude agreement more than directional accuracy, with greater use of relative-performance, sector, and broader-market evidence.The contribution summary describes broader evidence gathering through tools and distinct rankings for direction and magnitude.Source: §1 · p. 3

2. A Sandbox and Benchmark for Stock Price Forecasting

The section introduces a tool-using stock-forecasting sandbox and a chronological benchmark, then compares existing language models and quantitative baselines. Results show variation in overall scores, direction and magnitude, and performance across task groups and calendar periods.

2.1. A sandbox for market investigation

  • The sandbox lets models query market observations, interpret returned measurements, and submit one forecast using only information available at the decision-date close.Nine retrieval tools expose different evidence types, while lookback window, candle resolution, comparison group, and indicator parameters can vary.Source: §2.1 · p. 3, Table 1 · p. 4

2.2. Forecasting task and score

  • The task is to predict a stock’s log-return over one of eight horizons, and the terminal score rewards both correct direction and magnitude agreement.Wrong-sign and zero forecasts receive zero; evaluation selects tasks with |y| ≥ 0.5σ_H, yielding 240 of 398 test tasks.Source: §2.2 · p. 4

2.3. Benchmark construction and chronological split

  • The benchmark is outcome-selected, with 3,807 training tasks and 398 test tasks separated chronologically and by label-realization windows.The latest training label is realized December 29, 2023, before the first test decision on February 1, 2024.Source: §2.3 · p. 4, Table 2 · p. 5

2.4. How do existing models perform?

  • 49.84 for glm-5.3 leads existing systems, followed by claude-fable-5 at 44.48; momentum extrapolation scores 42.00, versus 20.94 for Qwen3-4B.The comparison covers fourteen systems before AURA-4B is added.Source: §2.4 · p. 5, Table 3 · p. 5
  • Directional accuracy alone does not determine joint ranking: the GBM reaches 85.8% direction but 37.9 magnitude agreement and a 32.56 score.Claude-fable-5 has higher directional accuracy than glm-5.3 (85.4 versus 79.2), but lower magnitude agreement (52.1 versus 63.0) and a lower total score.Source: Table 3 · p. 5, §2.4 · p. 6
  • GLM scores 44.16 on one-to-ten-day tasks and 57.40 on 21-to-126-day tasks; momentum extrapolation falls from 68.87 on easy tasks to 16.65 on hard tasks.Across easy, medium, and hard tasks, GLM scores 59.39, 47.42, and 44.54.Source: §2.4 · p. 6
  • GLM scores 60.94 in February 2024–July 2025 and 23.43 in August 2025–March 2026, while momentum extrapolation scores 41.99 and 42.01.The two calendar groups contain 169 and 71 scored tasks, respectively.Source: §2.4 · p. 6

3. Data Construction and Post-Training

The authors construct tool-use demonstrations from historical forecasting tasks, train Qwen3-4B on those demonstrations, and then optimize it with PPO using realized-return feedback. The stages move from learning demonstrated investigations to learning from the model’s own forecast episodes.

3.1. Training data construction

  • From 3,807 training tasks, the authors construct 3,631 usable demonstrations that connect market investigation with a numerical forecast, using historical outcomes as supervision.The same training task pool supplies episodes for PPO, where the student produces its own queries and forecasts.Source: §3.1 · p. 6

3.2. Supervised fine-tuning

  • SFT trains Qwen3-4B on assistant analysis, tool calls, and terminal forecasts, teaching the model to produce an investigation and answer in one sequence.The demonstrations are generated using realized training returns and task annotations in a teacher-only context.Source: §3.2 · pp. 6–7

3.3. Reinforcement learning from realized outcomes

  • PPO updates the SFT policy from self-generated forecast episodes using a terminal reward based on realized returns, with a KL penalty anchoring it to SFT.The task reward depends on the final prediction; there is no extra reward for more tool calls or longer analysis.Source: §3.3 · p. 7

4. Post-Training Results

SFT and PPO more than double the starting model’s forecast score, bringing AURA-4B near frontier systems on the benchmark. Its relative performance varies across horizons, difficulty groups, and calendar periods.

4.1. Improvements from SFT and PPO

  • 20.94 to 43.31: SFT and PPO raise the forecast score by 106.8%, with demonstration learning providing the larger initial gain.SFT raises the score to 37.94; PPO adds a further 14.2% relative to SFT.Source: §4.1 · p. 7, Table 4 · p. 7

4.2. Comparison with frontier models

  • 43.31 places AURA third among fifteen systems, above seven of nine frontier models and all four quantitative baselines.The two higher scores are 49.84 for glm-5.3 and 44.48 for claude-fable-5; momentum extrapolation scores 42.00.Source: §4.2 · p. 8

4.3. Results across horizons and difficulty levels

  • 48.91 versus GLM’s 44.16 on 1–10-day tasks, but 35.87 versus 57.40 on 21–126-day tasks.Figure 2 also shows AURA close to GLM on easy and medium tasks, while GLM leads on hard tasks; AURA beats momentum on medium and hard tasks.Source: §4.3 · p. 8, Figure 2 · p. 8

4.4. Results in the later evaluation period

  • 46.04 in the later period puts AURA first among fifteen systems, versus 42.17 in the earlier period.The starting model scores 17.41 in the later period and 22.43 in the earlier period.Source: §4.4 · pp. 8–9

5. What Does the Model Learn from Post-Training?

Post-training chiefly changes forecast magnitude agreement and market-investigation behavior. SFT expands tool use, PPO further shifts queries toward ranking and market context, and one episode illustrates these changes together.

5.1. Learning the magnitude of price movements

  • Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy moves from 62.9 to 65.4.The median absolute forecast grows from 0.012 to 0.042; a symmetric decomposition attributes approximately 94% of the Base-to-AURA score gain to the magnitude-associated term, measured on each model’s correct-sign predictions.Source: §5.1 · p. 9, Figure 3 · p. 9

5.2. Learning to gather broader market evidence

  • The share of ranking and market/sector aggregate queries rises from 16.2% at Base to 31.4% after SFT and 38.0% after PPO.Median calls and distinct retrieval tools rise from five and four at Base to eighteen and seven after SFT, then decline to sixteen and six after PPO.Source: §5.2 · p. 9, Table 5 · p. 10

5.3. A forecasting episode

  • On the DOW November 4, 2024 task, AURA makes sixteen calls and forecasts −0.0644 against a realized ten-day log return of −0.0796.All four models choose the same direction; AURA’s investigation includes candles, market and sector views, rankings, and indicators, and its forecast is closest in this example.Source: §5.3 · p. 10

6. Related Work

The related work spans economic accounts of investor behavior, evidence-gathering and time-series forecasting methods, financial agents and benchmarks, and outcome-feedback training. This paper combines market-tool interaction with post-training for continuous stock-return forecasts.

Investor behavior and price formation

  • Investor-behavior studies connect return chasing, fundamentals, entry, preferences, and beliefs to cross-sectional returns, motivating forecasts that use both recent stock performance and wider market context.The discussion also cites work relating catastrophe-bond prices to expected losses, issuer characteristics, and credit-market conditions.Source: §6 · p. 10

Information and tools for forecasting

  • Prior forecasting work includes agents that search for information, evidence-level aggregation, event retrieval, financial-news interpretation, and numerical time-series models.The cited sequence-model approaches include LLMTime, Time-LLM, Chronos, and Kronos; the paper studies a tool-using policy whose investigation and numerical output are both post-trained.Source: §6 · p. 10

Financial agents and benchmarks

  • Financial systems and benchmarks evaluate tasks from trading and quantitative reasoning to investment-research workflows; this sandbox instead sequences market-tool actions and scores a specified-horizon price forecast.It provides a common environment for comparing existing systems and training a forecasting model.Source: §6 · p. 11

Post-training with outcome feedback

  • Outcome-feedback research spans reasoning and interactive agents, event forecasting, and world-event forecasting; this setting pairs market-tool interaction with a continuous stock-return target.The paper constructs supervised investigation trajectories, then optimizes the student’s own episodes with PPO while analyzing predictions and tool use.Source: §6 · p. 11

7. Conclusion

The paper concludes that post-training improves Qwen3-4B’s stock-forecast scores on the benchmark while changing its market queries. The strongest numerical change is in magnitude agreement, and the evidence is limited to selected historical tasks.

7. Conclusion

  • SFT and PPO raise Qwen3-4B’s forecast score from 20.94 to 37.94 and 43.31, comparable to frontier models on the benchmark.After PPO, the median retrieval-tool count is six and ranking and market-context queries account for 38.0% of queries.Source: §7 · p. 11
  • The conclusion identifies magnitude agreement as the strongest numerical change, alongside increased use of ranking and market-context queries.The reported results concern outcome-supervised post-training on selected historical tasks.Source: §7 · p. 11

A. Task Construction and Scoring Details

This section details how tasks are selected and scored, describes the model-facing tools and prompt, and specifies the quantitative baselines.

A. Task Construction and Scoring Details

  • The benchmark retains direction-consistent tasks meeting outcome-quality criteria, so reported performance reflects a selected task distribution rather than a random sample of market opportunities.The collection includes 3,807 training tasks and 398 test tasks; construction annotations are not supplied to evaluated models.Source: A. Task Construction and Scoring Details · p. 15
  • The scoring mask selects returns with |y_j| ≥ 0.5σ_H, excludes nonfinite pairs, and reports joint score S, direction D, and conditional magnitude agreement M.M is averaged only over correct-direction predictions and is not probability calibration.Source: A. Task Construction and Scoring Details · p. 15
  • 21.11 points of the Base-to-trained score change are magnitude-associated, versus 1.24 direction-associated points; the decomposition is descriptive and rounded.The terms sum to slightly less than the displayed score change because the factors are rounded independently.Source: A. Task Construction and Scoring Details · p. 15

B. Tools and Prompts

  • The prompt restricts forecasts to point-in-time price-and-volume evidence and defines the target as the daily-VWAP log return over the requested horizon.The available actions include nine retrieval tools and a terminal prediction action.Source: B. Tools and Prompts · pp. 15–16, Table 6 · p. 16

C. Quantitative Baseline Details

  • Quantitative baselines use training-split price-and-volume information, without access to annotation rationales or difficulty labels.The baselines include historical-median, momentum-based, and GBM predictors.Source: C. Quantitative Baseline Details · p. 16

D. Forecast-Scale Diagnostic

The diagnostic rescales predictions to isolate output-scale sensitivity while keeping predicted directions fixed.

D. Forecast-Scale Diagnostic

  • 33.52 to 52.97: test-fitted positive rescaling raises gpt-5.6-sol’s score while leaving every predicted direction fixed.The selected factor is c = 3.65; this is an in-sample oracle diagnostic, not a separately fitted forecasting baseline.Source: D. Forecast-Scale Diagnostic · pp. 16–17

E. Complete Temporal Results

The temporal results compare systems across two calendar periods and report differences in the observed task mix.

E. Complete Temporal Results

  • Six-month tasks fell from 11.2% to 4.2%, and hard tasks from 30.2% to 23.9%, across the earlier and later calendar groups.The groups cover February 2024–July 2025 and August 2025–March 2026; the comparison uses tasks observed in each group.Source: E. Complete Temporal Results · p. 17, Table 7 · p. 17

F. The DOW Forecasting Example

The recorded DOW example contrasts models’ tool-call sequences and forecasts for a ten-day horizon, against a realized log return of −0.0796.

F. The DOW Forecasting Example

Related papers

Loading 2609.36914v1…