# Learning-loop review — 2026-10-08

1420 closed trades reviewed, 273 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# Learning review, 2026-10-08

## 1. Is overall expectancy holding?

Overall expectancy is **-0.30% per trade** (n=1420, win rate 45.6%, Wilson lower bound 43.0%, profit factor 1.26). Total P&L is +$2,787, so the dollar result and the percent result point in opposite directions. Position sizing explains this: larger positions are winning more than smaller ones.

The row count is large but the independent evidence is small:
- The pooled scoreboard mixes 6 accounts (paper and live) and 10 strategies.
- QSR writes several rows per decision. The power section puts its honest sample at 81 independent symbol-days.
- Paper accounts pay no fees and get optimistic fills.

So -0.30% is a description of the pool, not a measured edge.

Per account, QSR is negative everywhere except the tiny t212-investing sample (n=3):

| Account | QSR net%/t | n |
|---|--:|--:|
| paper-main | -0.25% | 631 |
| Paper02qsr | -0.47% | 62 |
| Paper03QSR | -0.65% | 95 |
| t212-isa | -0.66% | 145 |
| live-1 | -0.84% | 167 |

live-1 is a price-biased subsample, so don't read it as a clean signal.

## 2. Biggest drag

Only QSR and the pooled cuts have 20 or more trades. The largest drag is **RSI 30–40 at entry**:

| | n | Expectancy | Profit factor | P&L |
|---|--:|--:|--:|--:|
| RSI 30–40 | 447 | -2.15% | 0.48 | -$2,216 |
| RSI 40–50 | 493 | +0.87% | 2.10 | +$2,926 |

That is a gap of about 3.0pp. It nominally clears the 2.48pp noise floor, but I'm **not treating it as resolved**, for these reasons:
- The floor is for 5-day forward returns on 81 independent symbol-days. These buckets count rows, which are inflated by anchor/runner legs, multiple accounts, and 3–10 re-buys per ticker.
- The margin over the floor is thin, 0.5pp, before any correction for clustering.
- The bucket mixes strategies. RSI at entry is not a QSR gate; it comes from an unrelated indicator snapshot.
- The METRICS doc records six earlier "which entry condition predicts returns" findings that all failed re-examination.
- 226 trades have `unknown` RSI, so the cut is incomplete.

By strategy, the biggest dollar and percent drags are:
- **qsr**: -0.42%, n=1103. This is the dominant volume.
- **crypto-trend-v1**: -3.04% expectancy but PF 1.78, n=30. It is skewed by a few large winners.

## 3. Entry, exit, or regime problem?

- **Mostly an exit and payoff-structure problem, not a readable entry problem.**
  - Stop-loss exits: 710 trades, 2.7% win rate, -4.79% average, -$8,759.
  - Take-profit exits: +4.86% average, +$5,270.
  - Trailing-stop exits: +6.04% average, +$5,356.
  - Overall payoff ratio is 1.06 with a 45.6% win rate, so roughly break-even before costs.
- **MAE:** winners average -2.08% and losers -5.27%. Losers go straight to the stop and winners barely dip, which is consistent with an entry that is right or wrong quickly. That is not enough to say the entries are early.
- **MFE:** losers average +1.45%, which is only mildly positive. Some give back gains, but it isn't the dominant pattern.
- **Regime:** every labelled trade is SPY>MA200 and the rest are `unknown`. There is no bear-regime data, so no regime read is possible.

## 4. Proposed change

**NO CHANGE.**

- **Entry-signal changes are off the table.** Any RSI-band or similar threshold change rests on a gap that is too close to the noise floor. The projected revisit for a 1.0% edge is **2026-10-22**, and 2027-02-14 for 0.5%.
- **The exit parameters changed too recently to judge.** `hardSellPct` is now 8, `trailActivationPct` 8 and `stopAtrMult` 3. The realized take-profit exits (about 4.86%) mostly pre-date these values. Rule 4 says to judge changes only on trades made after them, and that sample doesn't exist yet. Changing another exit parameter now would also break the one-change-at-a-time rule.
- **Wait for a paired same-fill A/B instead.** The anchor-vs-runner comparison has 126 pairs and a noise floor of 1.23pp. It currently shows -0.90pp (95% CI -2.13 to +0.33), not significant. Re-check once the post-change pairs have accumulated, roughly 20–25 new pairs.

## 5. Open positions (context only)

- **Aging QSR rows with no time-stop:** HON at 57d (-0.94%), AZN at 28d (-3.93%), SYK at 23d (-3.70%), UNH at about 30d (-2.12%), and NFLX (-6.70%, the deepest underwater).
- **rsi2-v1:** INTC is -5.33% against a stop at 96.37.
- **Many winners sit at +5–7% without exiting.** OXY +5.0%, COST +5.6%, RELX +5.1%, COP +5.2%, SPOT +7.1% and INTU +6.6% are still below the 8% target and trail activation. They can give gains back before locking anything in.
- **Possible data anomaly:** one SHOP row shows TP 135.42 below SL 162.39. Check whether that is a short or a bad row.
- **Duplicates:** the open list repeats symbols because it pools anchor/runner legs and several accounts, so it overstates the number of distinct ideas.

## 6. QSR shadow comparison

The shadow window closed on 2026-08-23, so a fuller outcome backtest of legacy vs newlyA can be requested now.

- **Frequency:** the legacy method produced 48 would-buy events on 18 tickers. newlyA produced about 1,180 real entries (918 closed, 262 open) on 174 tickers. newlyA fires roughly 20–25 times more often and is far broader.
- **Overlap:** 14 of the 18 legacy tickers also appear in the newlyA set. The legacy set is largely a narrow subset, not a different set.
- **Overlapping tickers are mixed:** AEP, BTI, ENB, NKE, TRP, ITUB and VALE have poor realized results in the journal, while DDOG, IBKR, AZN, SU and EBAY do well. I'm not drawing a conclusion from this.
- **Possible follow-up for the backtest:** whether the extra breadth from newlyA is adding low-quality names. That is untested.

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

- **Overlap:** there are 37 overlapping open rows, but many are duplicate anchor/runner legs.
- **Entry bases:** most are within 1% of each other. The larger gaps are IBKR 2.8%, SLB 2.7% and WPM 2.9%.
- **The gaps go both ways:**
  - IBKR: t212-isa entered higher, which is worse.
  - SLB and WPM: t212-isa entered lower, which is better.
- **Nothing systematic yet:** there is no sign that pre-market-anchored entries run consistently worse. Targets sit at a similar percentage above each entry.
- **No outcome data:** both accounts show negative closed QSR net%/t (-0.66% and -0.84%), but they run different configs and sizing, so that is not a comparison of signal quality. A real comparison needs independently resolved closed trades on the same tickers and dates.