# Learning-loop review — 2026-09-25

1062 closed trades reviewed, 401 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# Learning review, 2026-09-25

## 1. Is overall expectancy holding?

Pooled expectancy is **+0.43% per trade** over 1,062 closed trades. The win rate is 52.1% (Wilson lower bound 49.1%), profit factor 1.70, total $5,228, and max drawdown $1,675. Positive, but not resolved:

- **The sample is smaller than n=1,062 suggests.** The same QSR signal is traded in several accounts at once, and each buy writes an anchor row and a runner row. The statistical-power section puts the honest QSR sample at 81 independent symbol-days.
- **QSR sits near break-even on its own arithmetic.** 816 trades, payoff 0.95, so the break-even win rate is about 51.3%. Observed is 53.9%, but the Wilson lower bound (50.5%) is below break-even, so zero expectancy can't be ruled out.
- **The result is almost all paper.** `paper-main` contributes $4,994 of the $5,228. The two real-money accounts are roughly flat combined: live-1 −$103 over 161 trades and t212-isa +$79 over 64 (net −$24). Paper fills are optimistic, and live-1 is a price-biased subsample.

## 2. Biggest drag

The biggest drag is **RSI 30–40 at entry**: n=288, 33.3% win rate, −1.84% expectancy, PF 0.62, −$1,034. It clears the 20-trade bar. No strategy bucket is a real drag. orb-v1 (n=68) is flat at −$67, and no ≥20-trade symbol is meaningfully negative.

**The gap does not count as resolved.**
- On paper the gap looks big. The rest of the journal averages about +1.27%, so the gap is about 3.1pp (3.6pp against RSI 40–50 at +1.79%). That exceeds the 2.48pp floor.
- But the floor assumes independent symbol-days, and these 288 rows are clustered. Several tickers have 0% win rates across 10–18 stacked rows (HWM, NKE, COHR, TRP, ENB, PDD, CMCSA), which points to a few stop-out episodes rather than a smooth RSI effect.
- The comparator "rest" also mixes in crypto and intraday strategies.
- A 3pp gap is far larger than any documented indicator effect. The METRICS doc says that makes noise more likely than discovery.

**Journal-integrity flags.** Several buckets have dollar and percent results that disagree, and some have negative hold times:
- `manual` exits: PF 0.18 but +$1,739.
- orb-retest-v1: 19% win rate but +$549, with an avg hold of −1.7d.
- qsr·UNH: 0% win rate, −3.24% expectancy, +$712.
- MARA and PLTR: avg holds of −6.5d and −2.1d.

Some rows probably have mismatched fills or timestamps. This should be checked before any bucket is trusted.

## 3. Entry, exit, or regime?

- **Exit: no clear problem.**
  - Losers' average MFE is +1.38%, under a third of the way to the ~4.7% average take-profit.
  - Take-profit exits average +4.67% and trailing-stop exits +5.97% (94.6% win rate), so targets and the runner are capturing what's on offer.
  - METRICS calls a *materially positive* loser MFE an exit problem, and 1.38% isn't that.
- **Entry: the likelier read, but unprovable.**
  - Losers' MAE is −5.13%, at or beyond the stop (stop-loss exits average −4.56%).
  - Winners' MAE is a mild −1.99%.
  - So the drag looks like names that keep falling to the stop with almost no bounce. Fixing that would be an entry-signal change, which the noise floor blocks.
- **Regime: cannot be assessed.** Every classified trade is SPY>MA200 (+0.12%), and the rest are "unknown" (+1.46%), which is missing data rather than a bear regime.

## 4. Proposal: **NO CHANGE**

RSI 30–40 is an entry-threshold bucket and its gap can't be treated as resolved, so the rule requires no change. At a 1.0% true edge the projected revisit date is **2026-10-09**. For realistic edges (0.5%) it is **2027-02-01**. I'm not inventing a substitute parameter.

Instead of a bucket change, I'd wait on the paired anchor-vs-runner test, which needs no out-predicting of the market:
- The current result is anchor −0.12% vs runner −1.20%, a difference of −1.08pp (CI −2.56 to +0.40) over 104 pairs.
- The sign has flipped from the earlier +0.40pp, which is what noise looks like.
- If the point estimate holds, about 195 pairs would make it significant, roughly 90 more.

Also wait on the journal-integrity fix above.

## 5. Open positions (context only)

- **Stacked repeat losers.**
  - BTI: about 8 rows at −2.5%, aged 44–46 days. Its 6 closed trades were all losses.
  - PDD: about 10 rows at −3.3%. Its 10 closed trades were all losses.
  - UBER: about 9 rows at −3.3%.
  - AEP: −4.2% at 46 days.
  - SYK is the deepest at −5.30% (10 days).
- **Rows above their listed take-profit.**
  - The take-profit is 1.05× entry (confirmed on ABT).
  - AMGN (+9.3%, TP 398) and KEYS (+12.9%, TP 336) sit well past their TPs, and EMR (+5.2%) is marginally past.
  - These may be armed runners, where the TP shown is nominal. But this is also the failure class of the 2026-09-21 t212 RKLB incident, so check the daily report's "Open issues" section.
- **Data anomalies.**
  - The SHOP row (qty 0.3264) has a stop above its take-profit (SL 144.91 / TP 135.42).
  - Several LSE/T212 rows show no P&L (BARC.L, NG.L, RIO.L, and others).
- **No time-stop.** BTI, AEP and others keep aging by design.

## 6. QSR shadow comparison

The shadow window is closed (2026-08-09 to 2026-08-23), so a fuller outcome backtest can be requested now.
- **Overlap:** the legacy method had 48 would-buy events across 18 tickers. The live newlyA method has entered 158 tickers. 14 of the legacy 18 overlap, and four (MO, MAR, MRVL, SNDK) were never bought by newlyA. newlyA is much broader, while legacy is narrow and mostly a subset of it.
- **Composition:** the overlap leans toward names that did poorly under newlyA. TRP, BTI, ENB and NKE went 0-for-all, and AEP, COHR, ITUB, TDG and VALE were also weak. Only DDOG, AZN, SU, EBAY and IBKR did well. This is qualitative only and covers few names.
- **Frequency:** rows versus events aren't comparable here, because rows include anchor+runner legs and multiple accounts.

## 7. t212-isa vs live-1 signal timing

There isn't enough data to say which account gives the better signal. t212-isa began real trading on 2026-09-11 and has no resolved-outcome comparison against live-1.
- **Overlap:** about 40 listed rows, many duplicated (anchor+runner), so roughly 25 distinct tickers.
- **Basis gaps are mostly small:** most are ≤0.7%.
- **Outliers:** DASH (4.5%), BSX (4.0%), VALE (3.5%) and NVS (1.4%).
- **No systematic bias:** t212 entered cheaper on DASH, BSX, CTAS and BP, but higher on VALE, NVS and ORLY. That doesn't match the "stale pre-market close" hypothesis.
- **Confounder:** the large gaps likely reflect different entry days. The table has no entry dates.
- **Targets:** each account's target is +5% off its own basis, so basis differences alone don't show a target-distance difference.

This becomes usable only after both accounts have enough closed QSR trades to compare.