Follow-up experiment design bundle — 2026-09-17
- Status: EXP-014 and EXP-018 run; EXP-013 and 015~017 pending
- User scope: a request to run after designing followed, so the highest-priority diagnostics were run first.
- EXP-014 and 018 produced manifests and ledgers; the remaining IDs fix their configuration hashes just before running.
- No future week is committed. This is the next research design, separate from the current work-week-6 results document.
1. What will be checked
| Run order | Reserved ID / design | Key question | Start condition |
|---|---|---|---|
| 1 | EXP-014 signal revalidation and return path | Does the RSI effect depend on the comparison group? When do returns arise? | Run completed · inconclusive |
| 2 | EXP-013 shared-capital RSI | Does it still translate into account returns with concurrent signals and cash constraints? | EXP-014 signal revalidation + shared-capital synthetic accounting validation |
| 3 | EXP-015 cross-sectional short-term reversal | Is there an effect from picking stocks that fell relatively hard on the same day? | Pre-audit complete · data_hold (no point-in-time universe) |
| Can start independently | EXP-016 ETF dual momentum | Is combining asset ranking with a trend filter better than simple holding? | ETF identification, raw prices, distributions, per-product costs |
| After data are secured | EXP-017 cross-sectional medium-term momentum | Does medium-term strength, excluding the most recent month, persist afterwards? | Point-in-time listing/delisting, corporate actions, total-return history |
| After each ledger is produced | EXP-018 map of effect conditions | Does the effect vary with liquidity, volatility, market state and signal concentration? | EXP-014 run · insufficient_data |
EXP-013 is kept because it is the number already used for shared capital in an earlier conversation. Because EXP-014's result was inconclusive, promotion of EXP-013 is on hold. The new strategy priority is EXP-015, and the comparison candidate with a different return structure is EXP-016.
2. Correcting the premise — why EXP-012 is not inherited as is
On 2026-09-17, scripts/backtest/run_connors_rsi2_exp012.py was read statically and the following was confirmed. No performance was recalculated and no existing ledger or verdict was modified.
| Finding | Limit on the earlier interpretation | Follow-up contract |
|---|---|---|
distinct_entries = sum(entry_days) | The same dates repeat across symbols. The earlier 253 cannot be interpreted as 253 independent days. | Record separately the size of the union of all symbols' entry_dates and the number of concurrent signals per day |
| When monthly matching fails, it widens to the whole year | This differs from the design's quarterly substitution. | The main comparison is fixed to the same-date universe; ex-post time matching is auxiliary |
| The 80% matching-rate gate is absent from the verdict code | It cannot be said that the whole sample gate is implemented. | Check denominators/numerators and missing reasons in the ledger |
| The standalone operation verdict checks only some conditions | Symbol/year concentration and all cost and period conditions were not reflected in the automatic verdict. | Each row of the verdict table shows the computed value, criterion, pass/fail and ledger path |
net_expectancy = mean net_pnl / initial_cash | +1.06% is a different value from the mean return on capital deployed per trade. | Name KRW expected P&L, per-trade net_return and P&L relative to initial capital separately |
| The bootstrap groups pairs by the month of the signal date | Whether it adequately reflects dependence from control days in other months and reused controls is unverified. | Same-date comparison, common-date blocks, including empty dates |
| Auxiliary holding periods, 3-month blocks, an ex-ante exposure benchmark and a full trade ledger were not implemented | Passing the earlier 131 tests does not guarantee the consistency of a new analysis. | State a synthetic validation list in each design |
Therefore +2.99%p is a stored exploratory result and is not used as confirmed signal evidence. "Portfolio construction is the bottleneck" is also a hypothesis to be tested. The new comparison groups and aggregation are split into EXP-014. The fact that the comparison method changed after seeing earlier results is not hidden, and the same criteria apply regardless of the sign of the result.
3. What was developed from the earlier proposal
- The EXP-013 base form is not chosen after seeing good holding periods in EXP-014. The RSI exit is fixed now.
- The absolute RSI criterion and cross-sectional ranking are separate experiments. An increase in trade count is not itself counted as an improvement.
- The top-100 liquidity universe is not fixed for all periods using 2010~2018 as a whole; instead, each decision date's 60-trading-day trading value builds the next month's universe. The survivorship bias of the current listing remains separately.
- Low exposure is an operational-suitability metric. A signal is not rejected for low exposure alone, and raising exposure does not guarantee returns.
- ETF and stock momentum start by checking total returns and point-in-time product classification. Missing products are not replaced on the spot with similar assets.
4. Common execution and verdict contract
Each document must be read together with the common contract. What is possible with current data alone is exploration with its limits flagged. Integer-share account results without actual trade prices and corporate-action linkage do not establish that a real 3,000,000 KRW account can be operated.
The completion criterion for this design is that hypotheses, controls, timing, accounting, matrices, failure conditions, required data and validation items are stated in each document. File hashes, point-in-time symbol lists, ETF codes and so on are fixed by mechanical rules in the next run preparation. Before that, they are not changed after viewing returns.
5. Scope of use of reference research
The relationship between short-term reversal and liquidity motivated the conditional analysis design. Dai et al., Reversals and the Returns to Liquidity Provision. Time-series trend and cross-sectional ranking are different hypotheses, so the ETF design separates the two components. Moskowitz·Ooi·Pedersen, Time Series Momentum. The 1st specifications are limited to reflect the problem that testing more specifications raises the chance of picking performance that arose by luck. Harvey, Backtesting.
The research above does not guarantee returns in Korean equities or under this account and cost setup. All thresholds and combinations below are this study's own preregistered proposals, not a contract copied from those papers.
Source and migration evidence
Source: docs/test-plans/2026-09-17-research-program.md
Original SHA-256: 1e93b650477dafcc06be51415618095fc7def8190163d9cad19bc5a71ab520d4
Raw data and account information are not part of the public package. The original results do not replace statistical or scientific revalidation. The English text is a translation; numbers, tables and charts are carried over unchanged.