# Learning-loop review — 2026-09-30

1210 closed trades reviewed, 348 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# Learning review, 2026-09-30

## 1. Is overall expectancy holding?
**Not distinguishable from zero.** Across 1,210 closed trades, expectancy is **+0.04%/trade**. Win rate is 48.4% (Wilson lower bound 45.6%), profit factor 1.72, payoff 1.08, and total P&L is +$5,443.

- **Dollar P&L and expectancy disagree.** A profit factor of 1.72 next to a 0.04% expectancy shows that P&L comes from a few large-dollar winners (manual exits, time stops), not from a steady per-trade edge.
- **The pooled number blends accounts.** On a net basis:
  - `paper-main` +$5,336
  - `Paper02qsr` +$117
  - `Paper03QSR` +$121
  - `t212-isa` +$0.33
  - `live-1` **−$139**, real money and the only account that pays fees
- **The row count overstates the evidence.** Most of the 1,210 rows are QSR anchor and runner legs, bought by several accounts on the same signals. The honest QSR sample is **81 independent symbol-days over 18 sessions**.
- **QSR is at break-even.** Its expectancy is −0.13% with payoff 0.96 and a 49.9% win rate.

## 2. Biggest drag bucket
- **Bucket:** RSI 30–40 at entry (n=346).
  - Win rate 31.8%, expectancy −2.00%, PF 0.64, −$1,025.
  - RSI 40–50 (n=440) sits at +1.14%, a gap of about 3.1pp.
- **Other candidates:** the MA50-distance buckets "−10..−5%" (n=290, −0.58%) and "−5..0%" (n=417, −0.55%). They are weaker and overlap heavily with the RSI one.
- **Does the gap clear the noise floor?** Nominally 3.1pp is above the 2.48pp floor, but **I don't count it as resolved**:
  - **The rows aren't independent.** 346 rows come from a small number of symbol-days. Repeated buys of one symbol across four accounts and two legs share one price path.
  - **The floor is computed at 81 independent observations.** Splitting 81 across RSI bands leaves about 20 per band. The standard error of a band-vs-band gap is then roughly 2.3pp, so a 3.1pp gap is about 1.3 standard errors.
  - **The loss is concentrated in a few names.** These symbols have 0% win rates: HWM (18 rows), NKE (18), TRP (14), ENB (11), CMCSA (10), PDD (10). Each is a handful of events repeated across legs and accounts.
- **Verdict:** this is a hypothesis, not a finding (anti-self-deception rule #5).

## 3. Entry, exit or regime?
- **Exit problem: unlikely.**
  - Average MFE of losers is only **+1.41%**, far short of the 5% take-profit. Losers never had a large gain to give back.
  - Trailing stops and take-profits exit at +6.1% and +4.7% on average, so exits are capturing what moves offer.
- **Stop too tight: unlikely.** Average MAE of losers is **−5.18%**, against winners at **−2.00%**. Losers went clearly against the position instead of being shaken out by noise, and stop-loss exits average −4.67%.
- **Regime: can't be tested.** Every row with a regime label is "SPY>MA200". The only other bucket is "unknown", which is the better one.
- **Most consistent read:** a selection or entry-timing pattern (buying names that keep falling), plus a payoff structure where the average stop (−4.67%) is about as large as the average target (+4.68%). Break-even then needs a win rate above 50%, and QSR sits at 49.9%. The sample can't confirm the entry reading.

## 4. Proposal
**Entry-signal and bucket-threshold changes: NO CHANGE.** That covers the RSI band, `nearMissZoneMaxPct` and any MA50 window. The supporting gap is within what 81 independent symbol-days can resolve. The table's projected revisit is **2026-10-14** for a 1.0% edge. Realistic edges of 0.5% or less don't resolve until about **2027-02-06** or later.

**One change that doesn't depend on out-predicting the market (a concentration cap):**
- **Param:** `qsr.maxBuysPerTickerPerMonth`
- **Current value:** 10
- **Proposed value:** 6
- **Why:**
  - **Re-buys stack exposure to one stop cluster.** Losing names absorb repeated buys (HWM 18 rows, NKE 18, TRP 14, all 0% win). Open books show the same stacking: PDD has 8+ open rows, BSX 9+, ABT and NFLX many.
  - **It frees capital.** The book holds 20–50-day-old losers with no time stop.
- **What this does not claim:** winners are also re-bought heavily (VST 27 rows, UMC 20, META 18), so a cap trims both sides. The case is lower single-name concentration and capital turnover, not higher expectancy.
- **Where to apply it:** the global default is 10. `paper-main` overrides it to 10, `live-1` to 8, `Paper03QSR` to 4 and `Paper02qsr` to 6. Apply it in `QSR_V1` or per account, your call.
- **How to judge it:** only on trades made after the change.

**Preferred longer-term test:** the anchor-vs-runner paired A/B. It is currently −1.12pp with a CI of −2.56 to +0.33 over 107 pairs, so not significant. Its noise floor is 1.44pp, about 1.7 times finer than bucket-splitting.

## 5. Open positions (context only)
- **Aging losers with no time stop:**
  - BTI about 50 days at −4.17%
  - HON 29–49 days
  - PDD 22 days at −2.95% across many rows
  - ABT, BKR (−5.94%), NUE (−5.40%) and APO (−5.62%), which sit near their stop levels
- **Skew toward the red:** most open QSR rows are underwater, so the closed-trade stats likely flatter the full book.
- **Inverted exit levels (SL above TP):** RCL (SL 261.16 / TP 250.66) and SHOP (SL 146.57 / TP 135.42) show the take-profit below the stop. That looks like bad data or a mis-keyed short, and should be checked.
- **Possible missing take-profit:** AMGN is +11.08% with TP 398.03. It is well past where a +5% target should have filled, so check whether its exit leg is live.
- **No P&L shown:** all crypto-trend positions and several fractional rows (GE, GEV, GS, NKE, CSCO, BN, URI, SPGI, BARC.L) show "–". That is expected for crypto, but it means they can't be monitored from this table.
- **rsi2-v1:** MARA is 25 shares at −5.26% after one day, and HOOD is −2.96%. This is early, with a single closed trade (n=1).

## 6. QSR shadow comparison
- **Status:** the shadow window closed on 2026-08-23. A fuller outcome backtest of the hypothetical legacy trades against the real newlyA outcomes **can now be requested**.
- **Breadth:**
  - Legacy produced 48 would-buy events across 18 tickers.
  - newlyA has made 742 closed and 321 open entries across 161 tickers.
  - The counts are in different units (events vs rows), so compare breadth, not raw counts.
- **Overlap:** 14 of the 18 legacy tickers also appear under newlyA, so legacy is essentially a narrow subset. The legacy-only tickers are MO, MAR, MRVL and SNDK.
- **Observation:** the overlapping set is mixed. AEP, TRP, BTI, ENB, NKE, ITUB, COHR and VALE are journal losers. EBAY, SU, DDOG, IBKR and AZN are winners.
- **Conclusion:** legacy isn't obviously better or worse. It is just far more selective. This is not evidence for point 4.

## 7. t212-isa vs live-1
- **Overlap:** 40 lines, which is about 20 distinct tickers duplicated across anchor and runner legs.
- **Entry basis:** most pairs are within 0.0–0.7% of each other.
- **Outliers:**

  | Ticker | t212-isa | live-1 | Gap |
  |---|---|---|---|
  | BSX | 44.20 | 45.985 | 4.0% |
  | MFG | 11.31 | 11.05 | 2.4% |
  | SLB | 51.535 | 50.345 | 2.4% |

- **Direction is mixed, not systematic.** t212 is lower on BSX and higher on MFG and SLB. That is weak support for a pre-market stale-price effect, and it doesn't point to one account's basis being consistently better.
- **Targets:** each account's target sits about 5% above its own entry, so target distance is the same in both.
- **Outcomes:** there aren't enough independently resolved trades to compare signal quality.
  - Net results are live-1 QSR −0.45%/trade (n=149) and t212 QSR −0.56%/trade (n=91).
  - The two accounts run different configs: t212 has trailing on and a $200 cap per buy, while live-1 has a 1-share cap and price-biased fills.
  - So that gap is not evidence either way.