Next experiment proposal — separating RSI(2)'s signal effect from operability

Next experiment proposal — separating RSI(2)'s signal effect from operability

• Written: 2026-09-17
• Status: top-level proposal — the detailed P1 contract is frozen in the EXP-012 document; not run
• Recommended single goal: decide whether daily RSI(2) produces after-cost returns beyond the effect of holding cash.
• EXP-012 is the planned ID for stage 1 below. Shared capital and new strategies have different contracts and get separate IDs.
• The sample, exposure and concentration figures below are proposed research-budget criteria, not proven statistical pass lines.

1. Judgments starting from existing results

Evidence: EXP-010 verdict, RSI(2) full-universe sweep, week 5 report, D-032~034.

Low exposure and small drawdowns alone do not prove the absence of a signal effect. Conversely, losing less than a falling buy-and-hold does not by itself show an effect. This experiment distinguishes these two interpretations. Values of MDD or CAGR simply divided by exposure are not used as statistical tests.

2. Priorities and stop conditions

3. P0 — minimum preparation to make the experiment interpretable

• State execution_model="v2", per-period capital isolation, and signal close → next valid trading day's open.
• Keep the code revision, configuration/data/universe hashes, runtime environment, seed and cost version in run.json. Prevent overwriting existing results and confirm ledger reproducibility on same-condition reruns.
• Check the handling of adjusted prices, corporate actions, trading halts and 0 volume. If the difference between integer share counts computed from adjusted prices and actual trade prices at the time cannot be resolved, the operability of a small account remains unconfirmed.
• Until actual CREON commissions are confirmed, write proxy, not actual. For taxes, state the existing per-product, per-period model, and confirm that frictionless means commissions, taxes and slippage are all 0.
• Flag the survivorship and selection bias from using the current listing and ex-post coverage conditions. Computing liquidity ranks from the train period alone does not remove this bias.

4. P1 — EXP-012: revalidate the same strategy but change the comparison

The detailed execution contract and dual verdict rules are fixed in 0. If the content below conflicts with the detailed document, the detailed document prevails.

Base contract to fix

• Reuse EXP-011's 30-symbol manifest and aggregate the existing 7 symbols / additional 23 symbols separately. The additional 23 were also already seen in the full-universe sweep, so they are not described as unobserved symbols.
• Base signal: close > SMA(200), enter on RSI(2) < 5, exit on RSI(2) > 70. No separate stop or time exit in the base form. The actual signal conditions are checked against the implementation and fixed at registration.
• An independent 3,000,000 KRW account per symbol, 95% of cash, integer quantities, no leverage. The results of the 30 accounts are not combined and reported as a 3,000,000 KRW portfolio return.
• 2010~2018, 2019~2022 and 2023~2026-08-26 are each run in isolation. The last period is reported as historical_recheck; only its mapping to the earlier locked_oos is kept.
• Explicitly define the cost scenarios: baseline costs, 2x commissions and slippage (taxes keep the statutory model), baseline + 5bp one-way, and fully frictionless — 4 cost scenarios.
• The 1st matrix is 30 symbols × 3 periods × 4 costs = 360 cells. Only the base form is judged. Additional runs of the risk-halt policy and neighbouring RSI 3/7 sensitivity are separate diagnostics, and no winner is picked.

A. Are returns right after the signal better than the control?

Hypothesis: even within the same uptrend, the rebound after an RSI(2) oversold signal is larger than on ordinary entry days.

1. The 1st measure is the after-cost return from the next valid trading day's open after the signal to the open 5 trading days later. 1/3/10 trading days are auxiliary diagnostics. This is event analysis, not an operating experiment that changes the existing strategy's exit rule.
2. Extract entry days above SMA(200) on the same symbol in the same month that do not use the RSI condition. Condition checks use only information confirmed before entry. The control is built with the same count and the same 5-day hold.
3. 1,000 draws with a fixed seed; controls that fall short are not forcibly filled but reported with the number of eligible events. Events extending past the end of each period are excluded identically on both sides by a preregistered rule.
4. Report the mean net return difference between the signal and control groups, the median and the P&L distribution. Uncertainty is computed by resampling with date blocks applied commonly to all symbols, accounting for simultaneous declines across symbols and overlapping holdings. Monthly blocks are the default and 3-month blocks a sensitivity.
5. This control is an ex-post statistical diagnostic, not a tradable benchmark. It is for comparing entry timing with matched holding periods and costs. Results before and after costs are both kept.

B. What does the actual existing strategy gain relative to cash?

For the account applying the actual RSI exit, compare cash (0 interest assumed), the same symbol's buy-and-hold, and buy-and-hold + cash invested at the average capital deployment of the past calibration period. The latter's weight is computed on 2010~2018 and applied fixed to 2019~2022, and computed on 2019~2022 and applied fixed to 2023~2026-08-26. The weight is not matched after looking at evaluation-period returns or exposure. Results of the calibration period itself are diagnostic and excluded from the comparison verdict. The limit of 0 interest on cash is also flagged.

Always read together: net expectancy, CAGR, MDD, trade count, number of actual entry days, share of position-holding days, average invested capital/NAV, and contribution by year and symbol. Results before and after excluding the single top symbol and year are also reported. Up and down regimes are defined by the prior day's market index vs SMA(200), not split by ex-post highs and lows.

Proposed criteria for passing to the next stage

• Sample: at least 100 completed trades and at least 50 distinct entry days in the recheck period for the additional 23 symbols. Even with many trades, trades clustered on the same date are not treated as independent samples.
• Signal: if the 1st 5-day event's mean after-cost control difference > 0 and the lower bound of the 95% block-resampling interval > 0, it counts as grounds for follow-up research. It is not replaced by a favourable result among the auxiliary metrics.
• Actual operation: base-form after-cost expectancy > 0, with the direction held under cost stress. Check whether the positive conclusion holds when one symbol or one year is excluded.
• Exposure: a median per-symbol holding-day share of 10% is proposed as an interim operating gate. Falling short blocks standalone promotion, but if the signal test is positive, P2 can check the shared-capital effect. 10% is not a statistical law, so it is fixed with its purpose before registration and not lowered after seeing results.
• Capital shortfalls and data anomalies are reported separately with cause and impact, not silently excluded to raise the pass rate.
• advance: meeting the conditions above moves to follow-up robustness research. It is not live-trading approval. hold: undecided because the sample is short or the confidence interval includes 0. rejected: with a sufficient sample, the upper bound of the control difference ≤ 0 and the existing strategy's after-cost expectancy is also ≤ 0. If only the cost stress fails, the verdicts on the signal and on the operating contract are separated.

5. P2 — shared-capital portfolio: is having few signals the same as capital sitting idle?

The current run_daily_portfolio, despite its name, runs a single symbol. Multi-symbol shared capital needs a new implementation.

• One total capital of 3,000,000 KRW, at most 5 symbols, at most 19% of pre-entry NAV per symbol, a 95% total deployment limit.
• Simultaneous signals are handled in ascending prior-day RSI order, ties by symbol code. The rules are fixed before running. Entries are processed after exits confirmed at the open, and cash from later intraday exits on the same day is not used for open orders. The budget is limited including integer quantities and commissions, and slot shortages, balance shortages and trading halts are each recorded.
• 3 and 10 slots are only sensitivity diagnostics. The base form is not changed to the slot count with the best results.
• A portfolio average invested capital/NAV of 20% is proposed as an interim minimum operating exposure. It is a different metric from the 10% per-symbol holding days, so the two are not conflated.
• Compare cash at the same total capital, a pre-fixed market ETF and an equal-weight hold of the 30 symbols. The equal weight reflects integer quantities and whether purchases are actually possible, and does not hide target weights that cannot be implemented.
• Validate cash preservation, simultaneous order sequence, limit breaches, integer quantities, corporate actions and separation of signal and fill dates with synthetic cases.
• CAGR 15% is shown separately as whether the existing business target is met. It is not itself used as a pass line for the existence of a signal.

6. P3 — forward paper observation after freezing the rules

• The start date is the first trading day after the final rules, code and universe are frozen. The writing date is not used as an automatic start date.
• Even if data after 2026-08-27 are newly received, past data from before the rules are frozen do not automatically become new OOS.
• The 1st review happens when both at least 6 months and 100 completed trades are met. Monthly checks focus on data and execution quality, and observation is not ended early because performance looks good. If the sample is still short at 12 months, it ends undecided and leaves a follow-up design.
• Record differences between expected fills and observable actual prices, missed orders and cost errors. Strategy changes midway are split into a new version and a new observation period.
• Unless the point-in-time universe and corporate-action issues are solved, the survivorship-bias limit of historical results remains.

7. If one more new strategy is validated

Next in line is ETF dual momentum. It is an untested candidate in the existing catalog, chosen for the development reason that it can use the current research base with daily data and monthly decisions . This does not mean its profitability has been confirmed.

The starting point of a separate draft is a simple contract: rank relatively by 12-month total return at month-end, hold cash if it is 0 or below, and switch at the next trading day's open. First check the ETF list, distributions, overlapping tracked assets and the common data period. Comparing only two similar domestic index ETFs would hardly test a cross-asset relative momentum hypothesis properly. If those data are not available, design the new collection scope first and put performance experiments on hold.

Minervini SEPA comes after that, since it adds financial-statement release timing and further data requirements. Expanding parameters of the rejected squeeze, picking the top-earning symbols from the RSI full universe and bulk collection of more 5-minute bars are low priority.

8. Execution order and completion outputs

1. After the P0 contract check, freeze the EXP-012 design. The current draft is not a run command.
2. Run the 360-cell base form and the event control, and break down by 7/23 symbols, year, exposure and cost.
3. Record advance / hold / rejected and whether to start P2. Even a negative result completes the experiment.
4. The completed files are configuration, manifest and run.json, signal/order/fill/cash ledgers, summary CSVs, equity, underwater, cost and exposure charts, control effects with confidence intervals, and the verdict document.
5. Tidying the official status of the 5-minute line and confirming actual commissions remain separate management items. This proposal does not automatically change past strategy verdicts or the next work week.

There is one completion question: "Do grounds remain to research RSI(2) further, and if so, which part — signal, shared capital or forward validation — needs to be checked?"