From Hindsight to ForesightAdapting Reusable Reasoning Structures from Recurring Financial Events

Abstract

• The problem. Financial events recur, while their drivers and economic conditions change.
• The approach. ReFoRM reuses past reasoning paths and checks each relationship against current evidence.
• The result. Forecasts become more accurate, with better probability estimates across five model families.

Recurring Financial Events

Earnings, inflation reports, and policy decisions return on a schedule. The benchmark groups comparable occurrences into families, so earlier events can inform later forecasts.

300

financial questions

40

recurring event families

100

forecast targets

The Recurring Financial Events Set groups comparable occurrences into 40 event families across five categories. A family fixes the entity or series, target quantity, unit, comparison rule, and resolution source, while the forecast cutoff, available evidence, numerical boundaries, and outcome change from one occurrence to the next. Each family recurs seven or eight times, which gives 300 questions: 190 with three classes and 110 binary. Questions are generated from immutable snapshots of 32 FRED series and SEC Company Facts for eight issuers, and class boundaries use only observations available before the cutoff.

Method

Start with resolved events, gather current evidence, test the historical relationships, and use the supported paths to forecast the next occurrence.

1. 01

   Build memory

   Keep the factors and relationships from resolved events.

2. 02

   Gather evidence

   Retrieve sources available before the new forecast cutoff.

3. 03

   Test relationships

   Preserve, reverse, contradict, or leave a relation unverified.

4. 04

   Forecast

   Use supported paths to estimate the new outcome.

Step through the Fed example. Hover over a factor or graph link to compare its evidence; on a touch screen, tap a factor.

1. Hindsight graph construction

   For each earlier occurrence in the same family that resolved before the forecast cutoff, ReFoRM builds a graph of factors, predictive relations, and supporting sources from documents available by its resolution. Each graph is reduced to complete root-to-target paths and the conditions under which they apply or fail. Resolved answers, realized values, and historical probabilities are removed.

2. Retrieving current evidence

   Factors in the hindsight graphs guide the search for evidence available before the cutoff, while historical relations, answers, and probabilities stay hidden. The retrieved documents form an evidence ledger that records the target operation, usable baseline, observed signals, supporting sources, and missing information.

3. Path retrieval and reconstruction

   Complete paths that match the ledger are retrieved intact, so each relation stays connected to its surrounding factors. Every relation is then assessed against current evidence instead of being assumed to hold.

   Preserved

   Current evidence supports the same effect.

   Reversed

   Current evidence supports the opposite effect.

   Contradicted

   The original relation fails without establishing its opposite.

   Unverified

   Evidence is insufficient to decide.

   Paths become active, contradicted, or unresolved according to these assessments. Reconstruction may remain partial, and current evidence always takes precedence over historical structure.

4. Forecasting with the current structure

   The reconstructed structure and the evidence ledger produce a reasoning trace before any option is chosen. Low, central, and high estimates in the target unit are mapped to the public answer boundaries to give the forecast distribution, and weak magnitude support leads to a less concentrated forecast. Historical probabilities are not reused, and no post-hoc calibration is applied.

Main Results

Explore the five-model average or select an individual model. Compare ReFoRM with each baseline across accuracy and probability quality.

Results by financial category

By category, ReFoRM has the lowest average Brier score and the best result in three of the five categories, and it improves on Direct Forecast in four. The gains span earnings, macroeconomics, markets, and monetary policy, while energy is nearly tied at 0.2186 against 0.2183.

Reasoning Analysis

ReFoRM produces reasoning with stronger evidence grounding and more specific explanations of how factors affect the forecast.

Reasoning ratings cover one Gemini 2.5 Flash Lite run and serve as supporting evidence.

Forecasting scores do not show whether a prediction follows from valid reasoning. After each target resolves, we build a reference graph from post-resolution documents, used only for analysis. Each reasoning trace is decomposed into atomic relations that a blinded judge checks against this graph. ReFoRM has the highest share of verified relations at 56.4%, compared with 50.0% for Factor Memory and 33.6% for Structure Memory, the two strongest baselines on this measure.

A separate LLM judge rated all 700 traces from one Gemini 2.5 Flash Lite run on five dimensions, without seeing the outcome or the forecasting scores. ReFoRM has the highest mean on every dimension, and none of its correct forecasts rests on a trace scored at or below 2.5. These scores are not ground truth and cover one model and one seed, so they serve as supporting evidence.

We also ask when reconstruction helps. Each forecast is grouped by how its retrieved relations were assessed, without using its prediction, outcome, or score. The largest Brier reduction over Direct Forecast comes from structures whose relations were reversed or contradicted. When a blinded judge confirms the assessment against the post-resolution graph, mixed and reversed or contradicted structures reduce Brier score by 0.039 and 0.032, while unconfirmed cases show no improvement. The gains depend more on reassessing relations than on carrying them forward unchanged.

Case Studies

See how relation reconstruction changes the predicted interval for U.S. CPI and clarifies the remaining uncertainty for Walmart revenue growth.

Citation

@article{jin2026hindsight,
  title={From Hindsight to Foresight: Adapting Reusable Reasoning Structures from Recurring Financial Events},
  author={Jin, Sangjin and Lee, Junhyeong and Han, Daehee and Kim, Kangmin and Park, Suhwan and Jeong, Seonghun and Park, Daekyung and Szpruch, Lukasz and Wang, Zhangyang and Choi, Chanyeol and others},
  journal={Available at SSRN 7475378},
  year={2026}
}