Study 103Calibration
Making uncertainty bands more honest on gap days
Widening ranges on non-earnings gap days improved coverage in the held-out test. More later outcomes fit inside the range, at the cost of a wider range.
Research / Public ledgerOpen access · Freely available
The question. The test.
The result.
We study market history to make the memory more useful for AI agents and people. Each study links its question, decision criteria, and available source documents. Read the findings, including the ideas that did not work.
Explore the evidence in your browser or through an AI agent. Each published finding stays tied to its own method and sample.
Read the latest research update → Short, dated explanations of published work. The study records below remain the source of truth.
New: The Chart Casebook Real charts, historical matches, and the differences that matter. Start with IONQ at noon and its 50 comparisons.
Research for agents How we plan to test evidence, clocks, citations and uncertainty in agent answers.
Alphaforge research
Stock notes & weekly views
Our dated interpretations of market states, changing expectations and observed responses. Sources, countercases and follow-ups included.
Paper positions
Research Model Portfolio
Simulated positions with 1-, 5- or 10-session horizons. A dated record of the research and its development, updated after the close.
Simulated. No real money invested in this model.Our philosophy
History informs the question.
Understand the present by studying how comparable situations developed in the past.
Research is free and expresses general views, without instructions to buy or sell.
16public studies
4passed their criteria
9did not pass
3other statuses
16 of 16 studies · Most recently decided first
Pass
Study 102Calibration
Same event does not always mean a better comparison
Matching the event type did not produce narrower ranges at equal coverage. The test also exposed weak coverage on event days, which Study 103 investigated.
Study 101Calibration
Closer analogs do not automatically deserve more weight
Weighting closer examples failed the primary test. Inverse-rank weighting narrowed the raw band but lost 5.2 percentage points of coverage; it did not establish narrower bands at equal coverage. Direction was not measured.
Study 100Memory
Similar charts narrow the spread among analog outcomes
Closer historical matches had less spread among their later outcomes. This does not establish a narrower forecast band for the stock being examined or predict its direction. Study 101 tested the weighting follow-up.
Study R62Tape
Seller exhaustion: decaying versus swelling volume into a low
One study arm was completed and its score remained sealed. The study is parked without a final result; the second arm has not been run.
Study 99Memory
Large price moves did not establish a reliable directional signal
Large price moves did not establish a reliable continuation signal under the study's tests. None of the 64 tested combinations passed the required checks.
Study 91Memory
Testing analog spread over one and ten observed bars
On the same records used by Study 90, analog spread was positively associated with absolute excess moves at one and ten observed bars. Both separately reported primaries passed. The secondary difference versus shape-only excluded zero only at ten bars. This was a horizon extension on a reused sample, not an independent replication or proof of an advantage at every horizon.
Study 90Memory
Historical analog spread was associated with later move size
Across 383 successful historical cases, wider analog spreads were associated with larger later absolute excess moves. A positive partial association remained after controlling for trailing volatility. The five-bar comparison with shape-only analogs was inconclusive: its reported interval included zero. This passed the original association test; it did not establish directional accuracy or better future forecasts.
Study RPO-1Fundamentals
Testing software backlog data as a comparison signal
Software backlog data did not pass the test for improving historical comparisons. The source records remain available; the signal was not added to ranking.
Study EV-0Events
Testing whether earnings context improves chart comparisons
Adding earnings context changed which examples were selected without improving the consistency of their later outcomes. This context remains visible but does not control ranking.
Study ER-003Edge refinery
Filing operating-cash-flow transition × prior-session short-flow crossing
The discovery test failed all required checks. A duplicate version of the same idea was also stopped, so the failed test could not be renamed and retried.
Study ER-002Edge refinery
Issuer buyback-blackout reopening
The source review found no dates that met the preset definition. The study stopped before inspecting returns or opening an outcome test.
Study ER-004Edge refinery
Legacy fundamental-factor deciles
The research family was retired because the historical evidence and implementation costs did not support continuing it.
Study ER-001Edge refinery
New buyback authorization × published days-to-cover
A promising headline result did not survive the checks. The evidence was weak, and much of the move appeared to precede the filing.
Study CAT-1Edge refinery
Catalyst impulse, balance and restart
The idea failed on the population selected before the test. It was not promoted or retuned after seeing the result.
Study PF-1Edge refinery
Priced financing: the first reclaim of the pricing level, long
The financing signal failed on its fixed 30-signal sample. Later versions also exhausted their tests, and the research line is closed.
Reading the record
A verdict has a specific meaning.
Pass is not a promise.
The study met its stated criteria on the evaluated sample. That is not proof of trading profitability, universal validity, or future performance.
Fail is a finding.
The hypothesis did not clear its decision criteria. The result remains visible, with the original question and the reasons it did not pass.
Not every study concludes.
Stopped work may end at a source or feasibility check. Parked work is paused. Read each study's record before interpreting its status.
Ledger updated . This is a curated public collection, not a claim that every internal experiment is published. How the research is conducted →