Rewarding Market Reasoning After the Market Answers

When news hits one company, its effect on related companies often arrives days later. Prices of suppliers, customers and competitors adjust slowly, because investors absorb information gradually and because real operations take time to transmit shocks. This delay is usually treated as something to model with numbers: find which stocks lead which, and by how many days. We propose to use it differently. The delay can serve as a verifier for open-ended language reasoning. A language model reads a news item on day $t$ and writes out a step-by-step argument, a chain of thought, for which other firms should be affected, in which direction, and roughly how many days later. The market then reports whether it was right over the following days. That gives a reward signal, noisy but objective, in a domain that has no answer key at the moment of reasoning.

The central question is whether reasoning skill learned on mathematics transfers to this setting. Training a model with reinforcement learning on verifiable math problems is known to improve coding and science performance. Nobody knows whether those gains survive contact with a stochastic environment, where the "proof" arrives late and is mostly noise. We ask that question in a form that a strong numerical lead-lag model cannot already answer. The test is whether math-trained reasoning produces propagation hypotheses that predict returns beyond what price-based lead-lag models already capture, especially for firms the news never mentions.

Core Claim

The claim is narrow and testable. Suppose a language model's propagation hypotheses are scored against realized returns after removing what a strong numerical lead-lag model already predicts. Then three things should hold if math-acquired reasoning transfers. First, a model whose reasoning was strengthened by on-policy reinforcement learning on mathematics should beat its own base model and a supervised-distilled counterpart on this residual signal. Second, the advantage should be concentrated in multi-hop cases, where the affected firm is two links away from the firm in the headline. Third, further on-policy training against the delayed, residualized market reward should increase that advantage without degrading general instruction following.

If none of this holds, that is also a useful result. It would mean that reasoning transfer from deterministic domains stops at stochastic, delay-governed ones, which is precisely what the transfer literature has not tested.

Where This Sits

Several existing systems already pair language models with lead-lag structure, but none of them asks a model to reason. LagLLM prompts a frozen LLM to judge whether candidate lead-lag edges are plausible in traffic and air-quality networks, then reorders input tokens so that leaders come before followers. It is the closest prior work. Its LLM is a static edge filter, though, it generates no arguments, and it never touches markets or reasoning transfer. CausalStock reduces each headline to five LLM-assigned scores and learns lagged causal graphs numerically. KHGT extracts knowledge-graph triplets offline. These systems motivate our design because they show text carries lead-lag information. They also constrain it: the LLM is kept away from the reasoning step, possibly for good reason, and we must show that letting it reason actually adds something.

On the numerical side, two results set the bar. DeltaLag learns which stocks lead a target, and at what lag, through cross-attention over past price-volume features. It shows that leadership is sparse, about 22 leaders per day in the S&P 500, and that it shifts across regimes. FININ shows that widening the news window to 20 days keeps improving prediction, which is evidence for slow information diffusion. Both supply the reference against which our residual reward is defined. Neither produces an auditable chain linking an event to a distant firm.

On the reasoning side, CPL trains abstract planning with tree search on math and reports gains on HumanEval and GPQA. Reasoning Vectors isolates the weight difference between an RL-trained model and its supervised parent and adds it to other models. UniReason shows that on-policy RL transfers while off-policy supervised distillation damages general ability, which it calls the SFT trap. We reuse these methods as our treatment arms. All three were evaluated only in deterministic domains. AlphaFin / Stock-Chain does apply chain-of-thought to stock prediction, but it treats each stock in isolation and trains with supervised fine-tuning. That makes it our representative of the supervised route that UniReason warns against. FinChain finds that multi-step financial reasoning collapses on harder problems, which it calls the "advanced cliff." This is a cautionary result that predicts multi-hop chains may fail, and our two-hop split tests it directly.

Research Program

Task. The unit of work is a news event about a firm $A$ on day $t$. The model writes a chain of thought and ends with a structured set of hypotheses. Each hypothesis has three parts: a firm $j$ drawn from the S&P 500 universe used by DeltaLag, a sign, and a lag bucket (1–2, 3–5, or 6–20 trading days). The model must name the intermediate link that justifies each hypothesis, for example a supplier relationship or a shared input cost. This makes the chain auditable, and it lets us tag every hypothesis as one-hop or two-hop.

Delayed, residualized reward. Let $u_{j,t,\tau}$ be firm $j$'s return over lag bucket $\tau$ after $t$. Subtract the market return and DeltaLag's out-of-sample prediction for that window. Then convert the result to a cross-sectional rank in $[-1,1]$ on that day. A hypothesis set $H$ with signs $s_h$ receives

$$R(H) = \frac{1}{|H|}\sum_{h \in H} s_h \, u_{j_h,t,\tau_h} \; - \; \lambda \, |H|,$$

where $\lambda$ is a small penalty on the number of hypotheses, so that naming every firm does not pay. Residualizing against DeltaLag means the model is rewarded only for information a strong numerical model misses. Ranks keep any single extreme return from dominating. This reward is the methodological core of the proposal. It turns the market's delay into a late-arriving check, and it decides which predictive information counts as new.

Treatment arms. All arms share one base family, Qwen2.5 at 7B, because Reasoning Vectors reports results on this family and the arithmetic requires a shared initialization. The arms are:

  • the base model;
  • the base plus a math-derived reasoning vector following Reasoning Vectors;
  • a math-RL-trained checkpoint;
  • a supervised-distilled counterpart trained on long math traces from a stronger teacher;
  • an AlphaFin-style supervised fine-tune on financial rationales.

Each arm is then optionally trained further with group-relative on-policy RL against $R$ on 2010–2020 events.

Contamination control. This is the largest validity threat. A model may simply remember how markets moved. We use two defenses. First, all firms are replaced with consistent pseudonyms, and the model receives a short anonymized profile per firm covering sector, major customers, and major suppliers. Second, the headline test window lies strictly after the base model's training cutoff. That requires collecting a fresh news feed, since TRNA ends in 2018. As a diagnostic, we ask models to name anonymized firms; if they can, the identities are leaking.

Baselines

The naive reference point is the sector-neighbor rule. Every firm in the same industry as $A$, plus its listed suppliers and customers, gets $A$'s news sign at the 1–2 day lag. It requires no model. If this rule matches the language models on residual rank correlation, then the reasoning adds nothing that a lookup table cannot provide.

The strongest numerical competitor is DeltaLag, which reports a Sharpe ratio of 2.1177 and annualized return of 0.2472 on the S&P 500 over its 2022–2023 test. We retrain it on our exact universe and split, with the same tuning budget as our largest hyperparameter search. Its predictions define the residual, so our models beat it only if their signal has positive rank correlation with what DeltaLag cannot explain. We additionally require that a long-short portfolio blending DeltaLag with our signal achieve a Sharpe ratio above DeltaLag alone in the post-cutoff window.

The strongest text competitor is a FININ-style encoder, run with the same 20-day news window and the same training data. FININ reports a daily Sharpe ratio of 1.772 on the S&P 500 index. It operates at index level, and no reported result was found on Lacuna for firm-level residual prediction with this architecture. We adapt it to output per-firm scores and residualize it the same way.

The supervised chain-of-thought competitor is Stock-Chain from AlphaFin, which reports 30.8% annualized return and 55.7% accuracy on its own data. We retrain it as LoRA fine-tuning of the same base on AlphaFin's rationales plus our training events, with matched compute.

Two ablations isolate the key ingredients. The first removes the chain of thought: the same RL-trained model emits hypotheses directly. This separates the value of deliberation from the value of the weights. The second replaces the residualized reward with a raw-return reward. This tests whether residualization is what keeps RL from relearning momentum that DeltaLag already captures.

Why not just…?

Why not just use DeltaLag and skip language entirely? On price data alone, DeltaLag and Hermes are very strong, so this alternative is a real threat. Its weakness is structural: a numerical model cannot move until some leader's price has already moved. A headline about a supplier strike is invisible to it until prices react. Residualizing against DeltaLag makes this testable. If the residual rank correlation of every language arm is indistinguishable from zero on the post-cutoff window, the reviewer is right, and we will report that.

Why not just embed the news with an encoder, as FININ does? FININ found that RoBERTa embeddings beat LLaMA-2 embeddings, so dense encoders are a serious competitor. They score firms that the text is about, however, and they have no mechanism for inferring that an unnamed customer two links away should react a week later. The decisive comparison is the two-hop split, restricted to firms not mentioned in any article that day. An encoder adapted to per-firm scores should have little signal there, and a reasoning model should not. If the encoder matches the reasoning model on that split, our claim fails.

Why not just prompt a large model without any math training? This tests whether transfer matters at all, as opposed to generic language-model knowledge. It is exactly the comparison between the base arm and the reasoning-vector and math-RL arms, all from the same initialization. A frontier model prompted zero-shot is a useful ceiling. It cannot be anonymized as convincingly, though, so we report it separately and treat it with suspicion.

Why not just fine-tune on financial rationales, as AlphaFin did? This is cheap, and AlphaFin reports strong returns. UniReason found that supervised distillation dropped IFEval from 69.2 to 42.3 while RL kept it at 70.0. We therefore measure IFEval and GSM8K before and after every arm. A supervised arm that matches RL on residual signal while preserving IFEval would refute the claim that the SFT trap applies here.

Decisive Checks

The primary metric is the rank correlation between each arm's hypothesis signal and DeltaLag-residualized returns. We compute it per lag bucket, split into one-hop and two-hop hypotheses, on the post-cutoff window, with day-clustered bootstrap intervals. Secondary metrics are three. The first is the blended-portfolio Sharpe ratio over DeltaLag alone. The second is lag-bucket accuracy, conditional on correct sign. The third is retention on IFEval, GSM8K and BBH, which tests whether RL against a noisy delayed reward damages general reasoning or, more interestingly, improves it. That would be transfer running in the reverse direction.

The proposal succeeds if the math-RL arms show positive residual correlation on unmentioned two-hop firms where the base arm, the encoder, and the sector-neighbor rule show none, and if delayed-reward RL widens that gap without IFEval loss. It fails cleanly if anonymization tests reveal leakage, if the sector-neighbor rule matches the language models, or if the no-chain-of-thought ablation performs as well as full reasoning. Each of these outcomes would be informative, and none depends on a backtest headline number alone.