# Learning-loop review — 2026-10-02 1340 closed trades reviewed, 294 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes). # Daily learning review, 2026-10-02 ## 1. Is overall expectancy holding? Not clearly. Pooled expectancy is **-0.28% per trade** over 1,340 closed trades. Win rate is 46.0% (Wilson lower bound 43.3%), profit factor 1.28, payoff 1.05, and total P&L is +$2,885 with a $1,675 max drawdown. Expectancy is negative while profit factor and total P&L are positive. This happens because % expectancy weights every trade equally, while dollars follow position size. Stop-loss exits at -4.75% average and take-profit/trailing exits at +4.7% to +6.0% are sized differently across trades. The row count is not the real sample. The power section says QSR has about **81 independent symbol-days** behind 334 scored rows. Anchor and runner legs, repeated buys, and several accounts all inflate n. The pooled number also hides a split by account, and fees matter: | Account | Net | Closed trades | |---|---|---| | paper-main (no fees, simulated fills) | +$3,021 | 853 | | Paper02qsr | +$11 | 62 | | Paper03QSR | +$125 | 95 | | live-1 (real) | -$188 | 204 | | t212-isa (real) | -$90 | 126 | QSR on the two real-money accounts is negative (live-1 -0.76% per trade, t212-isa -1.26% per trade). Paper-main's QSR is -0.20% per trade, but it is positive in dollars. The paper-main number likely overstates what the system nets in live trading, and the live samples are small and price-biased. ## 2. Biggest drag bucket Several buckets clear the 20-trade bar. The largest drag is **RSI 30–40 at entry**: n=414, 30.2% win rate, -2.19% expectancy, PF 0.46, **-$2,262**. - Outside this band the other 926 trades average about **+0.57%**, so the gap is roughly **2.8pp**. - That gap only barely exceeds the **2.48pp** noise floor. The floor was derived for independent symbol-day 5-day returns, and these 414 rows are heavily clustered, so the honest margin is below it. - The band is also mixed across strategies and accounts. Meanwhile RSI 40–50 earns +0.84%, which contradicts a dip-buying thesis and points to confounding by period or mix. I treat this as **unresolvable at the current sample**, not as the answer. The other candidates are weaker: - MA50 distance -10..-5% (n=341, -1.01%, -$806) has a smaller gap. - Exit reason "stop_loss" (n=677) is the mechanical flip side of the take-profits, not a bucket to fix. - Per-symbol zero-win clusters (NKE 19, PDD 19, HWM 18, TRP 14, BTI 13, ENB 11, AEP 11) look striking, but they have the most duplicated legs and are the least independent rows. ## 3. Entry, exit, or regime? - **Mostly an entry/selection pattern.** Losers' average MAE is -5.22% against winners' -2.02%, so losers go straight down to the stop. Winners' MAE of only -2% means winners don't need a wide stop. - **Not mainly an exit problem.** Losers' average MFE is +1.41%, which is positive but small. Trades that failed barely worked first, so tightening targets would not recover much. The 8% take-profit and trailing machinery is capturing +4.7% to +6.0% on the wins. - **Regime can't be assessed.** Every row with a known regime is SPY>MA200. There is no bear-regime sample, so no regime conclusion is possible. Expectancy turns negative because stops (-4.75% on 677 trades) outnumber targets (+4.72% on 343), and the payoff ratio is only about 1.05. Repeated buying of the same persistently falling names makes this worse. ## 4. One parameter change **Proposal: lower `qsr.maxBuysPerTickerPerWeek` from 3 to 2.** I am explicitly not proposing an entry-signal change. The RSI 30–40 gap (about 2.8pp, inflated by clustering) does not reliably clear the 2.48pp floor. - Revisit entry-signal questions at the **2026-10-16** projection (for a 1.0% edge, needing 124 symbol-days). - A realistic 0.5% edge needs 493 symbol-days, projected **2027-02-08**. The cap change does not depend on out-predicting the market. It is a position and concentration cap on stacking into one falling name. It is supported by the many tickers that lost on every trade (NKE, PDD, HWM, TRP, BTI, ENB, AEP) and by the open stacks in point 5 (BSX, NFLX). The evidence is observational and cluster-inflated, so treat this as risk control, not proven edge. Judge it only on trades after the change. Two caveats for applying it: - Several accounts override this value (paper-main 4, live-1 4, t212-isa 4, Paper02qsr 4, Paper03QSR 2, t212-investing 2). A change to the strategy default only reaches accounts without an override. To test it properly, apply it to **one account**, preferably a paper account such as paper-main, as a 4→2 override. - Applying it to one account keeps it a single change. ## 5. Open positions (context only) - **Stacked underwater QSR names.** - BSX has 5+ rows at -5.07%, held 15–24 days. - NFLX has several rows at **-12.41%**, held 3–14 days. - Also underwater: SPOT -5.47%, UBER -5.35%, APP -5.68%, SYK -4.23%. - These are the same repeated-ticker exposure that point 4 targets. - **Aged positions.** HON is +2.39% after about 51 days with no resolution. Several others have been held 22–31 days. QSR has no time-stop by design. - **Winning runners.** TXN +10.3%, COHR +9.3%, GLW +6.5%. - **rsi2-v1 on a speculative name.** MARA, 25 shares, is -6.10% after 3 days. Its stop (10.19) is probably still ahead, but it is the largest rsi2 drawdown. JNJ is -3.38%. - **Crypto-trend.** Roughly seven positions are aging 9–13 days. The table shows no live P&L for them. - **Data oddity worth checking.** - RCL shows SL 270.43 with TP 250.66. - SHOP shows SL 147.89 with TP 135.42. - For long positions the stop is above the target. This looks like a ratcheted stop on a stale or inverted TP, or a display problem. It is worth confirming that these rows are protected correctly. - URI also has an unusually tight stop/target (1048.61 / 1062.92). ## 6. QSR shadow comparison The shadow window closed 2026-08-23, so a **fuller outcome backtest of the legacy-method picks versus real newlyA outcomes can now be requested**. Observations: - **Volume.** The legacy method produced 48 would-buy events on 18 tickers. The live newlyA method produced 863 closed plus 270 open entries on 171 tickers. This overstates the true frequency gap, because the live figure counts rows (multiple legs, multiple accounts), not buy decisions. Even so, newlyA fires far more broadly. - **Overlap is high.** 14 of the 18 legacy tickers were also bought by the live method. The legacy picks are close to a subset of the live ones. - **The overlap names skew toward the journal's worst.** AEP, TRP, BTI, ENB, NKE and ITUB are all persistent losers in the journal (-0.84% to -5.5% expectancy, mostly 0% win rates). The legacy method may have been selecting the same recurring falling-knife names. This is observational only. ## 7. t212-isa vs live-1 signal timing - The 34 overlapping rows are about 17 distinct tickers, since most appear as anchor and runner pairs. Most differ by **0.0–0.9%** in entry basis. - Three differ by roughly 2.7–2.9% (IBKR, SLB, WPM). The direction is **not systematic**: t212-isa was higher on IBKR and BHP and lower on SLB, WPM, BP and UBER. - Both accounts' targets are a fixed percentage above their own entry (about +8%). The distance to target is therefore the same by construction, and a stale pre-market basis shifts only the absolute price level. - **There is not yet enough to say which gives the better signal.** The only comparable outcome data is closed trades, and the two accounts differ in config (trailing enabled, B+ buy, different caps), account size, and price-biased subsampling on live-1. Both are negative net so far, so any comparison would be premature. - Revisit once both accounts have a few dozen independently resolved QSR trades on the same tickers.