# Learning-loop review — 2026-10-09

1433 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, 2026-10-09

## 1. Is overall expectancy holding?

**No.** Expectancy is **-0.27% per trade** over 1,433 closed trades (win rate 45.7%, Wilson low 43.1%, profit factor 1.26, payoff 1.07). Total P&L is +$2,859 with a max drawdown of $1,755.

- **The sign is mixed.** The mean per-trade % is negative while dollar P&L is positive, so the large-sized trades are not the same ones as the frequent losers.
- **The row count overstates the sample.** QSR rows are clustered: anchor and runner legs, and several accounts trading the same signal. The power section says QSR has only about **81 independent symbol-days**.
- **The pooled number blends accounts and ignores fees.** The per-account table is the better read:
  - Paper accounts are net positive: paper-main +$2,961, Paper03QSR +$104, Paper02 +$11.
  - The live accounts are net negative: live-1 **-$194** (n=214) and t212-isa -$38 (n=150). t212-investing is +$8 on 4 trades, which is too few to read.
  - Paper fills are optimistic and fee-free, and the live accounts run different risk configs. live-1 is also a small, price-biased subsample.

## 2. Biggest drag

Many buckets clear 20 trades. I'm excluding the exit-reason buckets, because stop_loss, take_profit and trailing_stop are outcomes of a trade, not conditions you can act on.

The largest drag among the entry-condition buckets is **RSI 30-40 at entry**:

| Bucket | n | Expectancy | Profit factor | P&L |
|---|---|---|---|---|
| RSI 30-40 | 456 | -2.08% | 0.50 | -$2,165 |
| Everything else | 977 | about +0.58% (derived) | – | – |

The gap is about **2.65pp**, against a noise floor of **2.48pp**. That is barely over the floor, and the floor is defined for independent 5-day forward returns, not for clustered realized-P&L rows. After de-clustering, this bucket is **not resolvable**, so I'm not treating it as the answer.

- **The ordering is odd.** RSI 40-50 shows +0.91% and PF 2.11, which contradicts the "lower RSI is better" thesis. That points to confounds: strategy mix, exit configuration, and calendar clustering. It does not point to a real RSI effect.
- **Strategy buckets are flat or negative:**
  - qsr: n=1,115, -0.38%, but +$1,529.
  - orb-v1: n=68, about 0%, -$67. It is already at half size.
  - orb-retest-v1: n=44, 20.5% win rate, flat.
- **Per-account drag:** live-1 is net negative. Its worst strategy rows are crypto-trend-v1 (10% win rate, -7.53%) and crypto-scalp-v1 (29% win rate, -1.35% net after fees). Both are under 20 trades on this one account, so they are not actionable.

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

- **MAE:** winners average -2.09% and losers -5.29%. Losers go almost straight to the stop, which is about the stop distance. Winners are not showing deep early drawdown. I see no clear entry-timing signature.
- **MFE:** losers average **+1.46%**, so the typical loser was up before it reversed. That points mildly to an exit or give-back problem rather than a signal problem, but 1.46% is small against a 5-10% stop.
- **Exits:** 714 stop-outs (2.7% win rate, -4.81% each) cost $8,808. Take-profit and trailing-stop exits earned about $10,750 (364 at +4.93%, 191 at +6.04%). The system is stop-heavy: losses are about as large as wins, and the edge comes from hit rate.
- **Regime:** every bucket with a known regime is SPY>MA200. There is no bear-regime data, so a regime diagnosis can't be made.
- **Read:** an exit and payoff-shape issue, with a weak MFE signal. It is not an entry problem and not a resolvable regime problem.

## 4. Proposed change

**NO CHANGE.**

- **No entry-signal change.** The RSI 30-40 gap is within the clustered noise floor, so I'm not proposing one. The projected revisit date for a 1.0% edge is **2026-10-23** (124 symbol-days needed, 81 held). For 0.5% edges it is 2027-02-15.
- **What I'd do instead is a paired same-fill A/B.** QSR already produces paired data (126 complete anchor/runner pairs, noise floor 1.23pp). The current result is runner minus anchor of -0.90pp (CI -2.13 to +0.33) with the runner ahead in 41 of 126 pairs, so it is not significant.
- **Suggested test:** keep the pairing running and judge the runner against the anchor on net %. Revisit once pairs reach about 200 or the CI excludes zero. Because the MFE-losers number hints at give-back, a break-even-stop variant on one leg of each pair could be tested the same way. That is a design change, so it needs your approval and is not a parameter tweak.
- **Non-prediction levers** (sizing, caps) are fair game at any sample size. Nothing here shows a sizing parameter is clearly wrong, and cutting size would only shrink dollar losses without improving expectancy.

## 5. Open positions (context only)

- **Aging QSR rows with no time-stop:**
  - HON: 58.3d on one account (-0.46%) and 36-38d on others.
  - UNH: 30.3d.
  - BHP, AZN and PM: 29.3d.
  - FAST: 35d.
  - NTES: 35d.
  - UBER: 29-30d.
  - Many others are past 14d.
  - QSR deliberately has no time-stop, so these tie up capital with no scheduled exit.
- **Underwater now:**
  - NFLX: -8.33% (both legs, 10-21d).
  - INTC rsi2-v1: -7.76% after 3d, the deepest rsi2 position.
  - AZN: -3.68%.
  - SYK: -3.57%.
  - CRH: -2.9%.
  - BKNG: -2.74%.
  - APP: -2.32%.
- **Rows where SL looks at or above TP** (not a typo in the data, but worth checking):
  - MDB: SL 367.71 / TP 366.92.
  - SPOT: SL 509.18 / TP 508.57.
  - SHOP: SL 166.36 / TP 135.42.
  - INTU: SL 293.71 / TP 297.75.
  - MELI: SL 1793.59 / TP 1801.06.
  - These may be stops ratcheted by the trailing logic. Please confirm they still represent valid exit plans, especially on live accounts.
- **Fractional rows:** many Paper03/Paper02-style rows show a "–" unrealized P&L (e.g. FIX 0.0075), so the P&L display is blank for them.
- **Crypto:** the ADA/USD scalp is 0.3d old, and the AAVE trend position is 10.3d old.
- **Minor:** duplicated qty-1 pairs are anchor/runner legs, so the same position appears twice in the table.

## 6. QSR shadow comparison

The shadow window **closed** (2026-08-09 → 2026-08-23). A fuller outcome backtest of the legacy shadow versus real newlyA outcomes can now be requested.

- **Frequency:** the legacy method produced 48 would-buy events (38 buy-tier, 10 near-miss) on 18 tickers. The live newlyA method has made about 1,211 real entries (930 closed, 281 open) across **176 tickers**. The legacy method is far narrower.
- **Overlap:** 14 of the 18 legacy tickers also appear in the live set. Only MO, MAR, MRVL and SNDK are legacy-only. MO has live QSR trades in the journal but is absent from the post-switch overlap list, probably because those trades predate the switch.
- **Overlap names skew bad:** AEP, BTI, ENB, TRP, NKE and ITUB are among the journal's worst QSR symbols, each 0-45% win rate and net negative. DDOG, SU, EBAY and COHR look better or mixed. This is qualitative only and is not evidence for point 4.

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

- **Overlap:** 38 rows, about 19 tickers, each as an anchor/runner pair.
- **Size of the basis gap:** most are under 1%. The largest are WPM 2.9%, IBKR 2.8%, GLW 2.7% and SLB 2.7%.
- **Direction is mixed, not systematic.**
  - t212 entered higher on GLW, IBKR, WPM, NBIS and APP.
  - It entered lower on SLB, BHP, DASH, FANG and ORLY.
  - So I see no consistent pre-market or stale-price penalty on either side.
- **Targets move with the basis.** Because targets are computed from each account's own entry, distance-to-target comparisons are not independent evidence.
- **Verdict:** there is not enough data to say which account gives the better signal. Only open-position entry prices overlap, and neither account has independently resolved paired outcomes. Closed-trade results are confounded by config differences: t212 has a $200 per-buy cap and trailing enabled, while live-1 has a 1-share cap. Revisit when both accounts have at least 20 closed trades on the same tickers and dates.