harry/ForexBot2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Quant review 2026-09-26: alert cooldown, plain-language alerts, M15 grade-A SELL block, 1R printed targets, panel/breach fixes

Browse Source 
Live signals.db review (2026-09-04 → 09-25, 100 unique setups replayed on the
scans' stored candles): the stop-hunt claim is refuted (71% of stop-outs ran
≥ 1R past the stop, the same as the books' 70–72%); −25.1R of the −37.8R raw
live book were consecutive-bar re-entries the backtester's 5-bar cooldown never
takes; the cooldown view sits at the 3rd–5th percentile of the books' three-week
windows.

Delivery (book-neutral):
- telegram_bot: 5-bar alert cooldown per timeframe (_ALERT_COOLDOWN_BARS =
  backtest.cooldown_bars); setup block prints the trail first as the tested
  exit, TP1/TP2 as guides with their R, "Planned TP1 RR", and one validated-book
  line per setup class (agent_panel_base_rates.json); the informational 100-pip
  TP line is removed (owner request).
- AI block rewritten for a new trader: "AI check: the odds look better than /
  worse than / about usual", the chance "in 100" against the usual odds, one
  reason for and one against, no verdict word or confidence figure; rendered
  whenever the verdict carries a probability. agent_panel: persona prompts ban
  trading jargon (PANEL_PROMPT_VERSION v2.1-2026-09-26); the macro-source
  verdict now carries its persona's p_win (every live v2 row had
  panel_probability NULL). Dashboard and guides mirror the wording.
- telegram_bot: level-breach rescans skip the settlement-pause window (the
  20:45 M15 bar was re-logged at 22:01/22:06/22:11 every evening).

Signal (pre-registered rule-1 campaign, CUDA twin paired with gpu_backtest.py):
- M15 grade-A SELLs blocked (sell_min_quality ["A+","B"], M15 override 6):
  W2 813/+154.15R PF 1.53 vs 931/+146.14R PF 1.43, OOS +68.5 vs +60.67,
  W3 max DD 20.9 vs 19.3R. PASSED a–f.
- Printed target 1 floored at 1R in every generator (signals.tp_print_floor_rr),
  applied AFTER the min-RR and spread gates so no trade changes; mirrored in
  cuda_signal_kernels.py K2. Label-only re-pin of all four timeframes (only
  tp1/rr_planned columns move). A floor applied before the gate cost −24.6R and
  was rejected.
- FAILED: M15 min_rr_ratio 1.0 (trade-identical, −2.0R). Owner's call: M5
  tp_floor_rr 1.0 (+48.9R W2, max DD 10.1 → 15.9R), not shipped.

New pins: M5 W2 e599c8a8…, M15 W2 5b65ecf6…, M15 OOS 09b8a8e7…, H1 W3
a44fd19b…, H4 W3 f3587bae…; twin and reference MD5s equal on every window.
Base-rate table rebuilt from the new pins. Suite 558 passed.

Evidence: backtest/research/rr_and_sell_campaign_2026-09-26.md (plus the
2026-09-18 loss-anatomy and wrong-direction reports it builds on).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Harwinder Sekhon
2026-09-26 21:06:05 -04:00
parent 4f33480ce1
commit 1454ea1838
16 changed files with 1695 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backtest/research/loss_anatomy_2026-09-18.md
+165
View File
@@ -0,0 +1,165 @@
# Why the losing trades lost — two-year M5 and M15 books, 2026-09-18
Owner question: for every losing trade in the two-year books, was it the wrong direction, a late
entry, a target the market reversed before reaching, or a bad hour or day? Descriptive, in-book
statistics; nothing here is a decision (rule 1 needs a control-versus-candidate run).
## 1. Data
The validated two-year windows, 2024-09-05 → 2026-09-05, regenerated on the CUDA twin with the
decision-time features attached (`--export-features`), $0.65 round trip, factory configuration:
```
venv/bin/python backtest/cudaBasedBacktesting.py --pair XAUUSD --primary-tf 5min --start 2024-09-05 --end 2026-09-05 --spread 0.65 --export-features --export-csv w2_5min_feat.csv
venv/bin/python backtest/cudaBasedBacktesting.py --pair XAUUSD --primary-tf 15min --start 2024-09-05 --end 2026-09-05 --spread 0.65 --export-features --export-csv w2_15min_feat.csv
```
`Backend: CuPy GPU (NVIDIA GeForce RTX 4070 SUPER, cupy 14.0.1)`, 0 cache fallbacks, 781 and
931 trades, the pinned books (`49f8e602…`, `b4aa07e3…`) with the research columns appended. Every
stretch feature is signed in the trade's direction (positive = price had already moved the
trade's way), so buys and sells read alike. "R" is risk units: profit or loss divided by the
distance from entry to the initial stop. A loser is any trade with negative R.
| Book | Trades | Losers | Loss rate | Loss R | Win R | Net R | Full-stop losers | Trailing-stop losers |
|---|---|---|---|---|---|---|---|---|
| M5 | 781 | 368 | 47.1% | −278.9 | +424.9 | +146.1 | 247 (−257.8R) | 115 (−19.1R) |
| M15 | 931 | 368 | 39.5% | −341.0 | +487.1 | +146.1 | 324 (−336.4R) | 39 (−2.6R) |
Median loss −1.02R / −1.03R: a loser is almost always a full stop. The trailing exits' losses are
small by construction.
## 2. Price path of the losers (what the market did after entry)
| Path after entry | M5 losers | Share | M5 loss R | M15 losers | Share | M15 loss R |
|---|---|---|---|---|---|---|
| Never went our way (best excursion under 0.25R) | 153 | 42% | −157.9 (57%) | 162 | 44% | −166.9 (49%) |
| Shallow move, then reversal (0.25R to 0.75R) | 165 | 45% | −115.7 (41%) | 185 | 50% | −171.6 (50%) |
| Gave back a real advance (0.75R or more) | 50 | 14% | −5.3 (2%) | 21 | 6% | −2.5 (1%) |
- **Wrong from the start** is the largest single cause by R: over half the M5 loss and half the
M15 loss comes from trades whose best moment was under a quarter of the risk. Median hold 12
bars (M5) and 18 bars (M15): they were stopped fast.
- **Reversals after a genuine advance cost almost nothing** (1–2% of loss R): the trailing stop
arms at 0.6R (M5) / 0.7R (M15) and limits the give-back. On M5, 81 losers (22%) did reach the
full target-1 distance (median target-1 is only 0.51R on M5) and still closed red; they cost
−23.8R. On M15 that is 15 losers, −7.1R.
- Losers that reached at least half the target-1 distance: M5 157 (43%, −78.8R), M15 115 (31%,
−80.8R). So roughly a third to two fifths of losers "were winning at some point".
### Would a nearer hard target have helped? (all trades; exit at X when the best excursion reached X)
| Hard target | M5 losers rescued | M5 winners capped | M5 net | M15 losers rescued | M15 winners capped | M15 net |
|---|---|---|---|---|---|---|
| 0.50R | 142 (+118.0R) | 267 (−259.2R) | +45.4 | 102 (+118.5R) | 329 (−266.0R) | +58.6 |
| 0.75R | 50 (+42.8R) | 203 (−201.4R) | +72.1 | 21 (+18.2R) | 244 (−194.6R) | +97.5 |
| 1.00R | 3 (+4.8R) | 166 (−154.5R) | +97.3 | 2 (+2.4R) | 180 (−142.1R) | +128.5 |
| 1.50R | 0 | 94 (−91.5R) | +117.2 | 0 | 97 (−74.7R) | +143.5 |
| 2.00R | 0 | 53 (−56.2R) | +132.6 | 0 | 58 (−37.5R) | +152.2 |
| actual (trailing) | | | **+146.1** | | | **+146.1** |
A nearer target rescues losers but caps far more winners; the book's R lives in the right tail
(top 5% of trades = 94–96% of R, cycle 1). Only a 2R hard target on M15 edges the trailing exit
(+6R), which is the 2026-08-05 fixed-RR finding: it passes on 2024+ windows and is negative in
2021 and 2023 with double the drawdown. Not re-proposed.
## 3. Late entry: does chasing lose?
No. Loss rate falls as the entry gets "later"; the pullback-style entries are the worst bucket.
| Move over the previous 12 bars in the trade's direction | M5 trades | Loss rate | Avg R | M15 trades | Loss rate | Avg R |
|---|---|---|---|---|---|---|
| Against the last 12 bars (pullback entry) | 5 | 80% | −0.39 | 14 | 64% | −0.13 |
| 0 to 1 ATR | 60 | 50% | +0.18 | 43 | 44% | −0.06 |
| 1 to 2 ATR | 157 | 50% | +0.14 | 148 | 38% | +0.23 |
| 2 to 3 ATR | 180 | 42% | +0.28 | 288 | 41% | +0.19 |
| 3 ATR or more (chased) | 379 | 47% | +0.17 | 438 | 38% | +0.14 |
| Hourly stretch (distance above the 20-bar hourly mean) | M5 loss rate | M5 net R | M15 loss rate | M15 net R |
|---|---|---|---|---|
| Below the mean | 49% | +19.4 | 43% | +12.7 |
| 0 to 1 ATR | 50% | +21.9 | 45% | +25.0 |
| 1 to 2 ATR | 47% | +42.6 | 41% | +52.4 |
| 2 ATR or more (stretched) | 45% | +62.2 | 35% | +56.1 |
Same picture for the 48-bar move (M5: 4 ATR or more = 44% loss rate and +128.5R of the +146.1R),
the distance from the 21-bar average, consecutive closes and the exhaustion tags (two or more
tags: 45% / 40% loss rate, the best bucket on M5). This is the 2026-08-04 age/stretch verdict
again: the best trades are the stretched runs. "Late entry" is not a cause of losses here.
## 4. Wrong direction: what predicted the wrong side?
| Condition at entry | M5 trades | Loss rate | Net R | M15 trades | Loss rate | Net R |
|---|---|---|---|---|---|---|
| Timeframes agreed | 579 | 46% | +128.4 | 897 | 39% | +135.3 |
| Timeframes did not agree | 202 | 51% | +17.7 | 34 | 44% | +10.9 |
| With the trend | 475 | 45% | +93.5 | 596 | 40% | +68.1 |
| No trend, ranging | 169 | 49% | +25.7 | 120 | 37% | +47.0 |
| No trend, transitioning | 76 | 51% | +24.0 | 213 | 41% | +31.7 |
| No trend, mixed trending | 61 | 54% | +2.9 | (blocked on M15) | | |
| Confidence penalty 20 points or more | 94 | 55% | −3.3 | 11 | 64% | −0.4 |
| Buy | 683 | 47% | +137.3 | 635 | 37% | +146.6 |
| Sell | 98 | 49% | +8.8 | 296 | 46% | −0.5 |
- Disagreement between timeframes and a mixed regime raise the loss rate by about 5 points on M5
and remove most of the average R; the heavily penalised entries (20+ points) are net negative
on both books but are few.
- Indicator opposition does not predict the wrong side: the M5 bucket with **no** opposing
indicator loses most often (58%, 40 trades). Direction was not "obviously wrong" at entry.
- M15 shorts: 296 trades, 46% losers, −0.5R net. Every M15 R came from buys. This is the parked
cycle-2 candidate (`cycle_2026-09-10.md`); unchanged status.
## 5. When we lost
Entry hour is the open of the bar after the signal bar (Coordinated Universal Time). Cells under
30 trades are noise.
| M5 hour | Trades | Loss rate | Net R | | M15 hour | Trades | Loss rate | Net R |
|---|---|---|---|---|---|---|---|---|
| 08:00 | 62 | 55% | −0.7 | | 07:00 | 68 | 47% | −4.1 |
| 14:00 | 117 | 53% | +4.7 | | 09:00 | 75 | 47% | +3.8 |
| 17:00 | 35 | 49% | −4.0 | | 18:00 | 45 | 47% | −2.1 |
| 22:00 | 31 | 55% | +4.0 | | 04:00 | 28 | 43% | +1.9 |
| 20:00 | 42 | 52% | +30.1 | | 14:00 | 154 | 40% | +16.9 |
| 06:00 | 60 | 40% | +26.5 | | 12:00 | 68 | 31% | +33.8 |
The weakest hours are 08:00 and 17:00 on M5 (14:00 is the busiest hour and carries the most
loss R, −44.9R, at a 53% loss rate but still nets +4.7R) and 07:00, 09:00 and 18:00 on M15. No
hour is both large and clearly negative; 20:00 on M5 loses 52% of the time and is the second
best hour by R (+30.1R), which is the whole shape of this edge.
| Weekday | M5 trades | Loss rate | Net R | M15 trades | Loss rate | Net R |
|---|---|---|---|---|---|---|
| Monday | 153 | 45% | +17.6 | 169 | 39% | +28.2 |
| Tuesday | 159 | 48% | +21.3 | 212 | 33% | +50.4 |
| Wednesday | 176 | 45% | +46.6 | 189 | **51%** | **−4.0** |
| Thursday | 135 | 51% | +9.2 | 179 | 40% | +30.6 |
| Friday | 151 | 48% | +47.7 | 169 | 36% | +34.2 |
Wednesday is the one clear day effect: on M15 it loses more than half the time and is the only
negative weekday (−86.2R of loss R, the worst of any day); the worst M15 weekday-hour cells are
all Wednesday (14:00: 26 trades, 16 losers, −7.2R; 18:00: 13, 8, −4.7R; 09:00: 20, 14, −3.3R;
08:00: 16, 9, −3.0R). On M5 Wednesday is fine (+46.6R) and Thursday 20:00 is the worst cell (11
trades, 9 losers, −6.5R).
Calendar-day clusters are not the story: the ten worst days hold 11% (M5) and 9% (M15) of all
loss R, no single day cost more than −4.2R, and only 27 of 333 (M5) and 15 of 449 (M15) entry
days had three or more losers. Losers entered within two hours of the previous loser (re-entries
into the same move): M5 46 (−39.5R, 14% of loss R), M15 22 (−21.7R, 6%). By month, July is
negative on M15 in both years (2025-07 −7.5R at 57% losers, 2026-07 −8.6R at 58%) and M5 lost in
2024-Q4 and 2025-07/08.
## 6. Summary for the owner
| Question | Answer (both books unless stated) |
|---|---|
| Wrong direction? | Yes for about 43% of losers by count and about half of the loss R: they never moved a quarter of the risk our way and were stopped within 12–18 bars. Nothing at entry flagged them: indicator opposition does not predict it, timeframe disagreement and mixed regimes add only ~5 points of loss rate. |
| Late entry? | No. Loss rate falls as the prior move grows; the few pullback-style entries lose 64–80%. Stretched entries carry most of the R. (Confirms the 2026-08-04 verdict.) |
| Target too far, market reversed first? | Rarely the cause of a loss: reversals after a 0.75R advance cost 1–2% of loss R because the trailing stop limits them. About a third of losers were up at least half the target-1 distance before losing; a nearer hard target would rescue them but cap far more winners (net +45R to +133R against +146R). |
| Bad hours? | M5: 08:00 and 17:00 lose most often; 14:00 carries the most loss R but nets positive. M15: 07:00, 09:00 and 18:00. No large hour is clearly negative. |
| Bad days? | Wednesday on M15 (51% losers, −4.0R net, all four worst weekday-hour cells). Not a single-day-cluster problem on either book. |
None of this is a change proposal. A Wednesday or hour block on M15, or a shorter M15 SELL list,
would each be a rule-1 campaign with a control run, the OOS window and the refuted-hypothesis
list in mind (the 2026-07 M15 loss analysis rejected several hour variants). The most useful
follow-up, if any, is the parked M15 grade-A SELL block, which this table supports without
adding to it.