Large language models asked to forecast stock markets from a long window of monthly prices and earnings show a consistent failure. Told about reflexivity, the idea that investor beliefs and prices reinforce each other until a bubble bursts, they correctly see that a market is overvalued. They then exit years too early and miss the final rally (Reflexivity as Prompt). That study also found that the same theory prompt hurts GPT-5 with a 36-month window and helps it with a 60-month window. The usual reading is that the models reason about valuation but cannot time the market. We propose a different and testable explanation. Part of this behavior may come from where information sits in the prompt, not from what it says. Clinical sequence models have exactly this failure. They give spurious importance to whatever appears first in a patient record, a "primacy bias" that standard explanation tools cannot separate from real clinical signal (Interpreting Sepsis Prediction through Positional Explanation).
The proposal borrows that clinical tool, position-aware explanation, and adapts it to financial prompts. Each forecast is split into two parts. One is the share driven by the content of each dated observation. The other is the share driven by its position in the context. We then ask whether premature bearishness and the context-window reversal sit on the position side. If they do, a cheap position-randomized inference procedure should improve exit timing without retraining. If they do not, the "timing is hard" reading gains a control it currently lacks.
Core Claim
The positional explanation method PaX represents each input as a value paired with its position. It perturbs inputs with "POS-SWAP," which moves measurements to different positions while keeping their values. From these perturbations it computes separate importance scores for value and for position. In the sepsis setting it showed that Gender looked important only because it was recorded at time zero, while Age carried real signal (Kim et al.).
A naive transfer of this to price series fails, because in a time series, order is content. Swapping months destroys the trend the model is supposed to read. Our key move is to attach an explicit date to every row in the prompt (month, index level, trailing earnings), so the true chronology can be recovered from content alone. We then permute only the order in which rows are presented. A forecaster that reasons over the dated content should be nearly invariant to this permutation. A forecaster with presentation-order bias will not be. The same logic applies to the theory text: we move the reflexivity scaffold from the top of the prompt to just before the question.
The claim is falsifiable. Premature bearishness means the exit happens a measurable number of months before the index peak. If position-attributed importance accounts for a substantial part of that early exit, the hypothesis survives. If content attribution accounts for almost all of it, the hypothesis fails.
Research Program
The first component is the decomposition itself. Following PaX, we fit local surrogate models over three kinds of perturbation:
- Content perturbations, which replace a dated row's values with plausible alternatives.
- Presentation-order perturbations, which permute dated rows with values and dates fixed.
- Scaffold-placement perturbations, which move the theory block.
Two outputs are measured: the model's monthly position (long, flat, or short) and its stated cycle stage under Park's seven-stage mapping. Each yields a content score and a position score for every row and for the scaffold.
The second component is a diagnosis on the original episodes. Park's setting covers anonymized and normalized S&P 500 prices with Shiller trailing earnings, the Dot-com and Global Financial Crisis cycles, Conditions A through D, and 36- and 60-month windows. We test two specific questions. Does the bearish turn during the Dot-com mania load on early-presented rows or on the scaffold's position? Does the 36-month reversal for GPT-5 disappear once position effects are averaged out?
The third component is a synthetic control, which also addresses memorization. Even anonymized, normalized market history may be recognizable to frontier models. We therefore generate boom-bust series from a simple reflexive process in which a sentiment state feeds back into fundamentals. This is in the spirit of the speculative-wealth mechanism in Production Structure and Multiple Equilibria. We plant the peak at a known date and vary whether early-window information is truly predictive. This gives ground truth for both attribution faithfulness and timing.
The fourth component is the remedy. At inference time we average the model's decision over several random presentation orders of the same dated rows, plus two scaffold placements. This mirrors how PaX-informed attributions removed false positives without retraining.
How This Relates to Prior Work
Park supplies the setting and motivates the gap. It treats the LLM as an input-output forecaster and never asks what inside the prompt drives the bearish turn. Kim et al. supplies the method. We change its perturbation design, because their vital-sign rows tolerate position swaps while price rows need explicit dates to stay meaningful.
Two papers constrain how we read results. Drinkall et al. used kernel ridge regression and found far more causal links between news semantics and market shocks than linear models did: 19.5% of pairs versus 6.69%. This warns that a linear surrogate inside our attribution step may miss real content effects. We therefore also fit a kernel surrogate and report both. Vivaldi showed that moving arithmetic into a code sandbox removes hallucinated metrics, and that heavy orchestration degrades explanations from reasoning models. We treat a single code-tool call that computes valuation ratios as a separate control, not as our method.
Baselines
The naive reference point is buy-and-hold, together with Park's Condition A, which gives the model price data only. Park reports a combined Sharpe ratio of -0.024 for Gemini 3 Pro in this condition at 36 months.
The strongest existing method is Park's own theory scaffolding. Park reports the following Sharpe ratios:
- Gemini 3 Pro under Condition D at 36 months: +0.277 (combined).
- GPT-5 under Condition C at 60 months in the Global Financial Crisis: +0.674 (single cell).
We rerun these conditions with identical anonymization, windows, decoding settings and number of calls. Our position-averaged variant counts as beating them only under two conditions. First, it must reduce median months-early-at-exit in the Dot-com episode. Second, it must not lower combined Sharpe. The comparison uses matched total inference budget: averaging over orders uses as many calls as a self-consistency ensemble of the baseline prompt. That ensemble is itself a baseline.
The attribution baseline is feature-only LIME and SHAP over dated rows, which was the comparison point in Kim et al. There, PaX raised insertion AUC by up to 0.03 on MC-MED. No reported result was found on Lacuna for any attribution method on financial forecasting prompts, so we establish this number ourselves on the synthetic series, where ground truth exists.
The key ablation runs POS-SWAP with the date labels removed. If the date-labeled design is doing real work, removing dates should make the "position" and "content" scores collapse into each other.
Why not just…?
Why not just truncate the window and see whether bearishness goes away? Truncation removes positions and information at the same time. Park's 36-versus-60 contrast is already this experiment, and it cannot say which of the two mattered. We still run truncation, matched to our permutations. If truncation alone fixes timing as well as order-averaging does, we will report that the cheaper intervention suffices.
Why not just use SHAP on the rows? SHAP over rows treats a row's importance as a single number. That is precisely the confound PaX exposed, where Gender appeared important only because of its time-zero position (Kim et al.). The synthetic control, where the planted signal's position is known, tests directly whether row-level SHAP recovers it.
Why not accept that early exits are just valuation logic? This is the real threat. A simple rule that exits when the price-to-earnings ratio crosses a threshold would also leave the Dot-com rally early. We include that rule. If the LLM's exit dates track the rule closely and position scores are small, the hypothesis is rejected, and premature bearishness is a content-level property of valuation reasoning.
Decisive Checks
On synthetic series with known peaks, attributions must place weight on the planted signal rows. Faithfulness is measured by insertion AUC, where the planted rows are added back in attribution order. On the real episodes, we report three things: the fraction of variance in exit date explained by position scores versus content scores, the change in months-early-at-exit under order-averaging, and the change in Sharpe against Park's numbers above.
The outcome that would most change minds is a position-averaged GPT-5 at 36 months that no longer shows the reversal Park documented.
References
- Reflexivity as Prompt: Does Awareness of Self-Reinforcing Market Dynamics Improve LLMs as Financial Market Forecasters?
- Interpreting Sepsis Prediction through Positional Explanation
- Production Structure and Multiple Equilibria - A DSGE Perspective
- Stories that (are) Move(d by) Markets: A Causal Exploration of Market Shocks and Semantic Shifts across Different Partisan Groups
- A Multi-Agent Framework for Interpreting Multivariate Physiological Time Series