# Learning-loop review — 2026-10-05 1360 closed trades reviewed, 299 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes). ## 1. Is overall expectancy holding? Not clearly. Pooled expectancy is **-0.27% per trade** across 1,360 closed trades (46.0% win rate, PF 1.27, total P&L +$2,849). The % and $ figures disagree in sign because position sizes differ across accounts and strategies. The pooled number is not a trustworthy single sample, for three reasons: - **Account mix:** it blends accounts with different risk configs. Paper accounts are fee-free and net positive. paper-main is +$2,949, Paper03QSR +$150 and Paper02 +$11. - **Live accounts:** both live accounts with trades are net negative. live-1 is -$197.51 net over 205 trades and t212-isa is -$69.50 over 132. t212-investing has no closed trades yet. - **Honest sample size:** for QSR entry-signal questions it is 81 independent symbol-days, not 334 rows, and the noise floor is 2.48pp. The sample is large enough to say "no demonstrated edge on real money yet". It is not large enough to rank entry conditions. ## 2. Biggest drag bucket The largest drag that clears 20 trades is **RSI 30-40 at entry**: - n=423 - win rate 30.5% - expectancy -2.16% - PF 0.47 - total -$2,247 The rest of the journal is about +0.58%/trade, so the gap is roughly **2.7pp**. That is only just above the 2.48pp floor, and it is not safely resolvable: - The floor is calibrated on independent symbol-days. These 423 rows are heavily clustered, since the same stock and day produce several rows. - The band is confounded with strategy, account and calendar. RSI 40-50 does the opposite (+0.83%, n=479), which is more like a time-period effect than a stable RSI structure. - The MA50 -10..-5% band (n=346, -0.97%, -$776) is overlapping and partly the same trades. Treat RSI 30-40 as a hypothesis, not a result. Two other buckets are not causes. "stop_loss" exits (n=684, -4.76%) are an outcome, not an entry condition. "relvol<0.8" (n=952) is -0.70%/trade but only -$176 in total. ## 3. Entry, exit, or regime? Mostly entry, with a smaller exit give-back. Regime is not testable. - **Entry:** MAE of losers is -5.23%, close to the stop distance, while MAE of winners is only -2.05%. Winners rarely come near the stop, and losers go straight to it. Losers don't look like "the stop was too tight on a trade that would have recovered": winners never visit that depth. This matches the sweep note that widening the stop rescued nothing. - **Exit:** MFE of losers is +1.43%, a modest give-back. It is not the large, high-MFE-loser pattern that would point mainly at exits. Stop exits are about 2x take-profit exits (684 vs 346), at roughly symmetric sizes (-4.76% vs +4.75%). - **Regime:** every labelled trade is SPY>MA200 and the rest are "unknown", so a regime read isn't possible. ## 4. Proposed change **NO CHANGE.** The only candidate is an entry-threshold change (the RSI band, driven by the 30-40 bucket). Its supporting gap of about 2.7pp is at the noise floor given clustering, so I won't propose it. The statistical-power table puts the revisit at **2026-10-19** for a 1.0% edge, and **2027-02-11** for a realistic 0.5% edge. What to do instead: - **Paired A/B:** keep the anchor vs runner comparison running. Runner minus anchor is currently -0.96pp (95% CI -2.23 to +0.31, runner ahead in only 40 of 122 pairs), which is not significant. The noise floor on pairs is 1.27pp, so it needs more pairs before any exit-mechanic change. - **Non-predictive levers:** these are fair game at any sample size, but the live-account losses are small and mixed, so nothing here justifies a change yet. Examples are position caps, sizing, and the cost and slippage deltas between live and paper. ## 5. Open positions (context only) About 280 rows are open, overwhelmingly qsr, with no calendar time-stop. Several have been open a long time: HON 54d, BSX 27d, UNH 26d, and BHP, IBKR, EW and AZN at about 25d. Watch these: - **Deeply underwater:** NFLX at -11.99% (several rows). BSX is -5.4% and AZN -5.3%. MARA (rsi2-v1) is -6.9% and JNJ (rsi2-v1) -4.6%. - **Stops possibly not firing:** if I infer entry from the TP/SL levels, NFLX, AZN and JNJ look below their stop prices. TXN at +10.7% looks above its TP (280.10). That is an inference from partial data, but worth a quick check of the broker order state and `issues.json`, especially for the T212 accounts that poll their stops and take-profits. - **Crypto trend:** about 10 crypto-trend-v1 positions have been open 6-16 days with no P&L figure shown. - **Duplicates:** pairs of rows for the same symbol are the anchor and runner legs, not duplicates. ## 6. QSR shadow comparison The window is closed (2026-08-09 to 2026-08-23), so **a fuller outcome backtest can now be requested**. Qualitatively: - **Breadth:** the legacy method produced 48 would-buy events on 18 tickers. The live newlyA method has 877 closed and 281 open entries across 173 tickers, but that counts multiple legs and accounts, so frequency is not directly comparable. - **Overlap:** 14 of the 18 legacy tickers also appear in the newlyA entries. Several of them are repeat losers in this journal (AEP, TRP, BTI, ENB, NKE, ITUB), so the legacy method would have picked similar names rather than a better set. - **Legacy-only tickers:** MO, MAR, MRVL and SNDK have no newlyA entries since the switch. ## 7. t212-isa vs live-1 There is not yet enough to say which gives the better signal. The 36 overlapping rows are only about 18 distinct symbols, doubled by the anchor and runner legs, and none have outcomes yet. Entry bases are mostly within 1%. The larger gaps are WPM 2.9%, IBKR 2.8% and SLB 2.7%, and they go in both directions. t212-isa paid less on SLB and WPM and more on IBKR. There is no systematic bias from the pre-market-anchored entries. Closed QSR results are similar and both negative: t212-isa -0.99%/trade (n=129) and live-1 -0.80%/trade (n=165). The difference is unresolvable, and both samples carry the caveats above (live-1 has a price-biased subsample, and the accounts run different configs).