# Learning-loop review — 2026-10-06

1375 closed trades reviewed, 289 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# Learning review, 2026-10-06

## 1. Is overall expectancy holding?

Overall expectancy is **−0.21% per trade** (n=1375, win rate 46.3%, Wilson lower bound 43.7%, profit factor 1.28, total P&L +$2,999, max drawdown $1,698).

- **Mixed signals:** The unweighted per-trade % is negative while profit factor and dollar P&L are positive. Position sizes vary widely (QSR share caps, leverage cadence, per-account overrides), so a few larger winners outweigh many small losers.
- **Sample size:** The 1,375 rows are not 1,375 independent observations. The power section puts the honest QSR sample at about **81 independent symbol-days**, so this number cannot yet mean much.
- **Per account:**

| Account | Type | Net | Net %/t (QSR) |
|---|---|---|---|
| paper-main | paper | +$3,053 | −0.19% |
| Paper02qsr | paper | +$10.53 | −0.47% |
| Paper03QSR | paper | +$162 | −0.65% |
| live-1 | live | −$197.51 | −0.80% |
| t212-isa | live | −$36.69 | −0.66% |

- **Live vs paper:** Both live accounts are net-negative, and paper-main's positive total is carried mainly by crypto-scalp-v1 (+$821, a no-fee paper result) and crypto-trend-v1.
- **Caveats:** The accounts run different configs. live-1 is a price-biased subsample, and paper crypto pays no fees.

## 2. Biggest drag

Several buckets have at least 20 trades.

| Bucket | n | Expectancy | Profit factor | Total P&L |
|---|---|---|---|---|
| strategy `qsr` | 1072 | −0.39% | 1.17 | +$1,444 |
| RSI 30–40 at entry | 426 | −2.12% | 0.48 | −$2,201 |
| MA50 distance −10..−5% | 349 | −0.92% | 0.73 | −$752 |
| exit reason `stop_loss` | 686 | −4.75% | 0.13 | −$8,617 |

- **Biggest drag:** The largest dollar and expectancy drag among entry buckets is **RSI 30–40** (n=426, 31.0% win rate, PF 0.48). It is roughly 2.6pp worse than the rest (about +0.5% expectancy).
- **Can it be resolved?** Not reliably. The 2.48pp noise floor applies to 5-day forward returns on about 81 independent symbol-days. The gap is only marginally above that floor, and the 426 rows are heavily clustered, so the effective n is far smaller.
- **Confounds:** The bucket overlaps with MA50 distance and with strategy mix. The RSI 40–50 bucket at +0.89% also contains many different strategies. This bucket is a hypothesis, not a result.
- **Stop-loss:** The `stop_loss` row is the outcome rather than an entry condition, so it is not a bucket to tune. It does show the structure: 686 stops at −4.75% against 536 take-profit and trailing exits at about +4.8% to +6%. Payoff is about 1.0, so the system needs a win rate above roughly 50% to be positive before fees. QSR sits at 47.1% with a payoff of 0.98.

## 3. Entry, exit or regime problem?

- **Regime:** There is no regime signal. Every row with a known regime is SPY>MA200, and 349 rows are "unknown".
- **Entry (MAE):**
  - Winners average MAE −2.07%, so entries are not especially early.
  - Losers average MAE −5.22%, so losing trades go straight to the stop. That fits a signal that is simply wrong on those trades, not a stop sitting inside normal noise.
  - Whether stops are too tight cannot be told from these averages alone.
- **Exit (MFE):** Losers average **MFE +1.43%**. They briefly worked before stopping out. That is a mild exit or give-back signal, well short of the several-percent MFE that would clearly point to the exit.
- **Net read:** The picture is mostly an entry or selection problem on QSR, with a mild exit give-back component. The paired anchor-versus-runner test does not distinguish the two exit rules (anchor −0.55%, runner −1.51%, difference −0.96%, 95% CI −2.23 to +0.31, runner ahead in 40 of 122 pairs, not significant).

## 4. Proposed change

**NO CHANGE this cycle.**

- **RSI band:** RSI 30–40 is an entry-threshold change on a gap barely above the noise floor. The projected revisit is **2026-10-20 for a 1.0% edge** (124 symbol-days needed, 81 held). Realistic edges of 0.5% or less point to 2027-02-12 and beyond.
- **Exit mechanics:** The one thing that could be tested without out-predicting the market is stop or target mechanics as a **paired same-fill A/B**. For example, run `stopAtrMult` 3 against a wider value on twin legs of the same fill. The pairing resolves about 1.27pp versus 2.48pp for bucket-splitting.
- **Why not now:** There is no paired test running for stops yet, so it would have to be built before it counts as evidence. Changing it unpaired would not be.
- **Fees:** Live-1's loss includes about $6.79 of crypto fees.
- **What to wait for:** At least 2–3 more weeks of QSR rows (at 4.5 symbol-days per session, about 10 more sessions to reach the 1.0% row), a paired stop A/B, and per-account net%/t on each live account reaching about 20 independent trades.

## 5. Open positions (context only)

- **Underwater now:**
  - NFLX qsr is at −10.38% across multiple rows and accounts, held 7–18 days.
  - BSX is at −6.36% across about 7 rows, aged up to 28 days.
  - Also down: UBER −4.0%, AZN −4.0%, SYK −3.3%, rsi2-v1 MARA −8.35% (25 shares, 7 days) and JNJ −3.85%.
- **Aging:** HON has been held 33–55 days, and several QSR rows are more than 25 days old with no exit. With no calendar time-stop, QSR depends entirely on the stop or take-profit firing.
- **Suspicious exit plans:**
  - SHOP: SL 161.07 / TP 135.42 (TP below SL).
  - URI: SL 1066.45 / TP 1062.92 (TP below SL).
  - CIEN: SL 421.11 / TP 424.06 (a 0.7% band).
  - MELI: SL 1788.84 / TP 1801.06 (a 0.7% band).
  - Please check whether those stops and targets were built from stale or inconsistent prices. A TP below SL could fire immediately or never.
- **Crypto-trend:** It has 10+ open positions with no unrealized P&L shown, so mark-to-market is missing.

## 6. QSR shadow comparison

- **Window and frequency:** The window closed 2026-08-23, so a fuller outcome backtest of the shadow log against real newlyA outcomes can now be requested. The legacy method produced only 48 would-buy events on 18 tickers. The live newlyA method made 887 closed and 272 open entries on 173 tickers, about 25 times more often.
- **Overlap:** 14 of the 18 legacy tickers also appear in the real entries. Most of those overlaps are tickers the real journal lost on:

| Outcome | Tickers (real-journal net) |
|---|---|
| Losers | AEP (−$299), BTI (−$392), ENB (−$288), TRP (−$264), NKE (−$206), COHR (−$229), ITUB (−$60) |
| Winners | SU, IBKR, AZN, EBAY |

- **Takeaway:** This is observational only, and it suggests the legacy list is concentrated in names that already underperformed here.

## 7. t212-isa vs live-1

- **Overlap:** There are 36 overlapping rows. Because of anchor/runner legs, these are about 18 unique tickers.
- **Entry basis gaps:** Most are under 1%. The three larger gaps are about 2.7% to 2.9%:

| Ticker | t212-isa entry | live-1 entry |
|---|---|---|
| IBKR | 88.64 | 86.25 |
| SLB | 49.045 | 50.345 |
| WPM | 133.83 | 137.71 |

- **No systematic direction:** t212 is higher on IBKR and lower on SLB and WPM. Targets are a fixed percentage above each entry, so target distance matches by construction.
- **Verdict:** There is not enough to say which account has the better signal. No matched closed outcomes are available for these pairs, and the pre-market anchoring effect is not visibly systematic here.