Building
by
Yongjae Lee
Jacob Chanyeol Choi
Our ICAIF 2026 paper shows that an AI forecaster does better when it remembers why past releases moved, and re-tests those reasons against new evidence each time.
What we found
Memory helped when it was re-tested. Across 100 recurring financial events, ReFoRM picked the correct outcome 55.0% of the time, versus 49.7% for a forecaster that starts fresh each time.
Memory reused as-is did not help. Reusing an earlier event's map unchanged scored 49.2%, no better than starting fresh. The largest gains came when ReFoRM found that a past relationship had broken or reversed.
The result held across models. ReFoRM had the best probability scores on all five language models tested.
AI can learn from past financial events. Knowing what drove the last few releases gives a forecaster a place to start. The next call depends on which of those forces are still at work.
It’s November 2026, and you’re preparing a forecast for the month’s U.S. Consumer Price Index (CPI), due December 10. Prices rose 0.6% in September and 0.4% in October. Would you pencil in another 0.4% for November?
Consider what drove those two readings. According to the U.S. Bureau of Labor Statistics, shelter costs rose in both months. Gasoline alone accounted for more than half the September increase in CPI, but the rise in gasoline prices slowed from 9.1% to 1.0% in October.
Now look at what’s changed. The national average price of regular gasoline fell from $4.153 a gallon on November 2 to $3.995 on November 9, according to the U.S. Energy Information Administration. If that decline lasts, gasoline could ease some of the pressure from shelter. That gives you a concrete assumption to revisit as the rest of November’s prices come in.
Rising shelter costs and cheaper gasoline could pull November CPI in opposite directions. Past readings show seasonally adjusted monthly changes. This opening is an illustration, not model output.
How ReFoRM Builds a Memory for the Next Release
A company reports earnings every quarter. Oil prices, payrolls, credit spreads and policy rates can all be tracked from one month to the next. We built ReFoRM, short for Recurrent-event Forecasting with Reasoning Memory, to learn from that repetition. The study covered 40 recurring series across corporate earnings, energy and commodities, macroeconomic indicators, markets and credit, and monetary policy.
ReFoRM builds a map for each past event from reports explaining its outcome. The paper calls these maps reasoning structures. For example, suppose a natural gas report says a cold spell increased demand for heating. Temperature and heating demand are the two factors. The link is how colder weather drove demand higher. Past answers and probabilities are removed before a map is reused; only the factors, the links between them and their sources are kept.
Across five events in the same series, the factor names form a list of things to look up. Each link stays with its original event and the evidence and conditions that supported it. For the next forecast, ReFoRM uses the list to find current reports, then checks which past links still fit.
200 past events informed 100 forecasts. Matching shapes mark the same factor. Numbers identify events.
In the two examples below, we compare ReFoRM with Direct Forecast. This baseline starts fresh with each forecast question, uses it to find reports available by the forecast date, and assigns probabilities from that evidence alone. ReFoRM also checks which links from earlier releases still hold.
Example 1: Was May’s CPI a Good Guide to June?
For June 2025 inflation, Direct Forecast picked the low range and ReFoRM picked the middle. On June 30, May’s 0.1% CPI rise was the latest reading. Direct Forecast expected much the same for June. Its estimate of 0.099% fell in the low range, below 0.146%. All six comparison methods listed in Table 1 favored that range.
ReFoRM had maps of five earlier CPI readings spanning January 2023 to December 2024. Consider two links from those maps. In June 2024, persistent housing costs had slowed the decline in inflation. In December 2024, higher food and energy prices had pushed headline CPI up.
By June 2025, housing costs were still putting upward pressure on CPI, so the housing link held. The food and energy signals were mixed, leaving their combined effect open. ReFoRM forecast a pickup to about 0.20% and put a 60% probability on the middle range.
June CPI rose 0.254% in the study data, inside the middle range. Direct Forecast gave the low range 85%. The Bureau of Labor Statistics initially reported +0.3%, a rounded figure; ranges were scored on the unrounded change.
Example 2: How Likely Was a Sharp Slowdown at Walmart?
For the quarter ending January 2026, both models expected little change in Walmart’s year-over-year net sales growth. They differed most on the risk of a sharp slowdown. Direct Forecast weighed strong sales against pressure on consumers and gave a sharp slowdown and a sharp acceleration 25% each, leaving 50% in the middle.
ReFoRM checked what had sustained sales before. In the October 2023 quarter, grocery and online demand had lifted total sales. In the October 2024 quarter, online growth had strengthened the holiday outlook. By November 2025, U.S. shoppers were buying more groceries by volume, global online sales were up 27%, and Walmart had raised its full-year net sales outlook.
ReFoRM judged direction and size separately. With both sources of demand still growing, the direction was clear, so it put just 5% on a sharp slowdown. How much a 27% online gain would lift sales companywide was less clear, because online was only part of Walmart’s business. ReFoRM moved most of the freed probability to the middle (65%) and only slightly raised the odds of a sharp acceleration, to 30%. Net sales growth eased from 5.81% to 5.64%. Using unrounded figures, the change was −0.175 percentage points, inside the middle range.
Both models picked the middle range (−1.385 to below +1.349 percentage points). Sales growth eased by 0.175 points.
What Changed Across 100 Forecasts?
For CPI, checking past links changed which range the model picked. For Walmart, both models picked the middle, but ReFoRM saw much less risk of a sharp slowdown. Did that help across all 100 forecasts? Averaged across five language models and three runs per model, ReFoRM picked the correct outcome 55.0% of the time, compared with 49.7% for Direct Forecast. Its probabilities were also closer to the outcomes on average, with a Brier score of 0.217 versus 0.231.
How much did the links add once the model knew which factors to look up? Factor Memory searched by those factors and received the same current articles as ReFoRM. It reached 50.9% accuracy. ReFoRM could also check the relationships among those factors, and its accuracy was 4.1 percentage points higher.
The average did not hide a split. ReFoRM had the best Brier score and log loss on all five models, and the highest accuracy on four of the five.
Accuracy scores the top call. The Brier score also accounts for the odds on every other outcome. A lower Brier score means the probabilities were closer to what happened. Changes in accuracy are in percentage points (pp) relative to Direct Forecast.
The biggest gains came when ReFoRM found that a past relationship no longer held, or had reversed. Reusing an earlier map without checking it, as Structure Memory did, scored 49.2%, no better than starting fresh.
Did the reasoning hold up after the release? We built a fresh map from reports about each outcome, and an AI reviewer compared it with the links each forecaster had relied on, with method names hidden. 56.4% of ReFoRM’s links matched, versus 50.0% for Factor Memory. A match required agreement on the factors, the direction of the effect, the outcome being forecast and the time horizon. A second review found that none of ReFoRM’s correct forecasts rested on weak reasoning, compared with 24% for Direct Forecast.
What This Does and Does Not Show
The study shows that, for recurring financial events, a memory of past relationships that is re-tested against current evidence can improve both the top call and the quality of the odds. It does not show a forecaster that is usually right: 55% accuracy is a reminder that these calls stay hard.
The questions asked which of three ranges, or which of two outcomes, a release would fall into. They did not ask for returns. The benchmark covers 40 series and 100 forecasts, not a live investment process. The reasoning reviews used an AI judge, and the weak-reasoning comparison covers one model and one run, so treat it as supporting evidence.
What This Means for Research Teams
Post-release notes can become an asset. Most teams rebuild their view from scratch each time a company reports or a data series prints. What drove the last few outcomes lives in one analyst’s head and leaves when coverage changes hands. Keeping the factors, the links between them and their sources as a structure lets the next forecast start from what the team already learned.
Memory pays off when it is questioned. Reusing last time’s reasoning unchanged did no better than starting fresh. The largest gains came when the forecaster noticed that a past relationship had broken. In a research process, that means asking at every preview which parts of the last thesis still hold and which do not, and writing both down.
Reasoning becomes auditable. Every link carries its source, so a reviewer can see which past relationships a forecast kept and which it dropped. That makes it possible to tell a forecast that was right for good reasons from one that was lucky, which is what a portfolio manager needs before relying on the number.
The edge sits in the memory, not the model. The gains appeared on all five models tested, including small, low-cost ones. What the system remembered, and how it re-tested that memory, mattered more than which model did the reasoning.
What to Keep for the Next Release
For a series you cover regularly, try ending your post-release note with a question you can reopen next time. After October’s CPI report, that note might read, “Shelter costs are still rising. If gasoline gets cheaper, how much of that pressure could it offset?” As November’s data arrive, you know which prices to follow and what would make you change your call.
Paper
From Hindsight to Foresight: Adapting Reusable Reasoning Structures from Recurring Financial Events
Authors | Sangjin Jin¹,*, Junhyeong Lee¹,*, Daehee Han²,*, Kangmin Kim¹, Suhwan Park¹, Seonghun Jeong², Daekyung Park³, Lukasz Szpruch⁴, Zhangyang Wang⁵, Chanyeol Choi³, Youngjun Kwak²,†, and Yongjae Lee¹,³,† |
|---|---|
Affiliations | ¹ UNIST / ² KakaoBank / ³ LinqAlpha / ⁴ University of Edinburgh / ⁵ The University of Texas at Austin. * Equal contribution. † Corresponding authors. |
Venue | 7th ACM International Conference on AI in Finance (ICAIF 2026) |
Presented | November 14 to 17, 2026, Milan, Italy. Poster session on November 17 (afternoon). |