EXP-018 — A map of the conditions under which the effect occurs

EXP-018 — A map of the conditions under which the effect occurs

• Status: partially run · insufficient_data (based on the EXP-014 event ledger, 2026-09-17)
• Contract: common contract
• Question: does the reversal effect differ between broad market declines and single-stock declines? In which liquidity and volatility buckets is it concentrated?
• Scope: first the validated event ledger of EXP-014, then the EXP-013/015 account ledgers. EXP-012's +2.99%p is not fixed as the dependent variable before the validity checks.

1st hypothesis and fixed classification

The 1st hypothesis is that 'even for the same RSI signal, the 5-day excess effect differs between broad market oversold days and non-concentrated days'. Crossing the preregistered buckets below to build dozens of strategies is not done.

The liquidity rank is not market capitalization, so the names large/mid/small cap are not used. Quantile boundaries are set on the same-day eligible set, and ties are broken by symbol code. The market close and RSI are known after day t's close and can be used for t+1 orders. Ex-post highs/lows and next-day moves are not put into the classification.

Returns and dependence

• For each event, the same-date 5-day excess effect produced in EXP-014 is used as is. The control group is not re-chosen after seeing the condition to enlarge the effect.
• 1st statistic: concentrated-day daily-mean effect − non-concentrated-day daily-mean effect. Each side needs signal days ≥30, events ≥100 and observed months ≥12.
• The CI of the difference is computed by a common-date bootstrap. Cells with too few samples stay NA/hold.
• Other axes are reported in tables of per-bucket mean, median, CI, costs and counts. Small-sample extremes are not hidden.
• Additional tests are limited to 3: market direction difference, liquidity top-bottom difference and volatility top-bottom difference. These 3 each use a two-sided Bonferroni α=0.05/3, and results before and after correction are shown together.
• Even if effects differ by condition over the full period, they are not immediately promoted to an operating rule of 'when to trade'. Whether a direction found in history_A also holds in history_B is compared as preregistered, but both periods are labelled exploratory.

Concentration risk to check under shared capital

• Link day-t signal counts, next-day actual fill counts, slot/cash/liquidity rejection counts and the actual capital deployment rate.
• Publish the number of concurrent positions and per-symbol loss contributions on the worst 5 loss days.
• Separately identify 'many signals but too few slots', 'entering together and losing together' and 'signals themselves are rare'.
• Sector analysis runs only as a supplementary appendix when a point-in-time sector classification exists. Missing sector, financial or market-cap data are not estimated and filled in from symbol names or the current classification.

Scope and decision

• The base table is 2+2+3+3=10 univariate buckets. Cross-searching all combinations is excluded.
• The conclusion is one of stable_pattern / unstable / insufficient_data. Stable requires sufficient samples in both periods, the same direction in history_A and history_B, a 1st CI excluding 0 and the same direction under cost stress.
• Even if stable, existing strategy filters are not modified automatically. For example, testing 'avoid signal-concentrated days' requires registering a separate new contract, including a future observation plan.
• New capital allocation, stop or leverage strategies are not outputs of this diagnostic.

Future outputs and validation

Produce univariate effect tables, a market signal-rate time series, tables of concurrent entries and cash/slot shortfalls, worst-day contributions, and tables of missing data, minimum samples and multiple-testing adjustments. Validate input-time consistency across symbols on the same date, exclusion of indicator-missing symbols from the denominator, the inclusion direction at the 10% boundary, unions of different dates, reproducible tie classification, and that replacing future prices does not change the current classification.