EXP-014 — RSI signal revalidation and the path where returns arise

EXP-014 — RSI signal revalidation and the path where returns arise

• Status: run completed · inconclusive (2026-09-17)
• Prerequisite contract: common contract
• Order: before EXP-013. A new ID is used because the comparison group and aggregation differ from EXP-012.

Questions and hypotheses

H1: among stocks in an uptrend on the same date, those with RSI(2)<5 earn more than the rest of the uptrend set over the next 5 trading days. H2: where does that difference arise — intraday on the first day, in overnight gaps, or over the following days? H1 is the 1st test and H2 is explanatory. H2 is not used to choose an exit day and change the EXP-013 base form.

Fixing the sample

• Use the EXP-011 30-symbol manifest as is. The existing 7 symbols are auxiliary; the other 23 are the 1st signal set.
• Base signal: day t close > SMA200, Wilder RSI2<5. Indicator initialization and warm-up are checked against the existing implementation.
• 1st events are a non-overlapping sample that skips signals on the same symbol for the next 10 KRX trading days after its first signal. This is determined only from earlier signal dates, not by looking at future prices or exit results.
• On the same date, at least 5 of the 30 symbols must be above SMA200, have valid indicators/raw prices and be tradable on day t. This is used as the equal-weight comparison set B_t. RSI signal symbols are included in B_t too. The claim is 'additional predictive power relative to the uptrend set', not a difference from a set with no signals at all.
• Comparisons excluding RSI signal symbols from B_t, common stocks only, and the full sample of raw repeated signals are auxiliary diagnostics. Flag that excluding ETFs/preferred shares changes the sample composition. They do not replace the main comparison.

Timing and return decomposition

• Entry at t+1 open O1 and exit at t+6 open O6 are fixed as '5 trading days open-to-open'. The KRX calendar is used. The next CSV row is not assumed to be the next trading day.
• Base path O1→C1→O2→C2→…→C5→O6. The sum of each segment's log-return must equal log(O6/O1). Simple returns are not summed.
• The 1-, 3-, 5- and 10-day paths are O1→O2/O4/O6/O11 respectively. On the same path, costs are charged once each at entry and exit. The common sample comparing all paths is limited to signal dates where 11 opens can be secured before the end of the period. The tail with only 5 days is a separate boundary diagnostic and is not added to the 1st sample.
• Intraday MAE/MFE are measured separately using the lows/highs from O1 to C5. The order of the high and low cannot be known from daily bars, so these values are not used to build optimal stop/take-profit fills.
• If the new entry day is not tradable, the fill fails. If halted while held, the scheduled exit waits and the actual exit delay is recorded. Fixed-time marked returns and delayed realized returns are shown side by side. Events are not deleted because of a future halt. If still open at the end of the period, it is in a marked/exit-pending state and not shown as a confirmed return.

Statistics and revalidation

For each signal symbol, compute the 5-day after-cost return difference vs the same-date B_t, average within the signal date, then weight signal dates equally. The 1st cell is history_B × generalization group × proxy_slip2 × non-overlapping sample. Even when the same B_t is used several times, per-symbol observations are not counted as independent samples.

A common 21-trading-day bootstrap and a 63-day sensitivity are applied. The full calendar including empty dates is resampled together. For dates where the effect cannot be computed because of trading halts or unconfirmed exits, the missing ratio and the unrealized P&L of those symbols are published, and promotion does not rely on a complete-case analysis alone that makes main-comparison losses disappear.

An auxiliary procedure to check the specificity of the historical control is also designed. Among eligible symbols on the same date, 1,000 label-placebo draws sample the same number as the signal set uniformly. This is not proof of a tradable portfolio or of a causal effect; it is a diagnostic with market co-movement held fixed. Its seeds are separated from the common seed.

Matrix

• Main table: 3 periods × 4 costs × 4 holding periods = 48 effect summaries. Only the 5-day cell above is judged.
• Auxiliary: the repeated-signal sample, control 7, the common-stock sample and non-signal B_t in separate tables; choosing the best combination per period is forbidden.
• Matching/samples are reused identically across costs; only trade prices and costs change.

Verdict and branches

• Proposed sample criteria: 100 non-overlapping signals in the 1st set, 50 union signal dates across all symbols, 12 or more months with signals, valid comparison days / eligible signal days of 80% or more. Short of this is hold.
• signal_supported: 5-day net excess effect >0, lower bound of the 21-day block CI >0, positive 63-day block point estimate, positive point estimates at both stress costs. If same-date common shocks or corporate-action defects remain, data_hold.
• signal_not_supported: upper bound of the 1st CI ≤0 with a sufficient sample. Means the added value of the RSI signal could not be confirmed.
• Otherwise inconclusive. Positive auxiliary 1/3/10-day results do not overturn it.
• If supported, start the main EXP-013 experiment. If inconclusive, accounting and concurrent-signal diagnostics are possible, but no profitability promotion. If not_supported, stop RSI-extension performance experiments and move to EXP-015/016.

Synthetic validation before implementation and output design

• O1/O6/O11 dates, holidays/halts/period boundaries, warm-up signals, and that future data changes do not alter day-t selection.
• Whether simultaneous signals on two symbols on a common date count as '1 distinct date'; whether the non-overlap rule removes repeated signals.
• Price path sums agree, no double cost charging, NA when there are no signals, reuse of the common same-date B_t.
• Trade/signal/candidate ledgers, exclusion reasons, daily paired effects, return paths, MAE/MFE and sample distributions, bootstrap settings, CIs and per-period/per-cost tables will be kept. Results are currently unmeasured.