SniperGold_ML/docs/P2_3_CONTEXT_WINDOW.md

# P2.3 — RUNTIME CONTEXT WINDOW (minimum history requirement)

Status: **RESOLVED — verified 0/0/0/0 mismatch (14850 rows)**  
Date: 2026-08-21  
Source of truth: **runtime deployment window = 700-bar M15 cache**  
(A 600-bar LOOKBACK alone is technically insufficient — see §3)

---

## 1. Runtime facts (from EA code + verification)

```text
M15 cache  : maxBars = max(InpMaxBars=700, 2*SWING_LEN + LOOKBACK) = 700
             (the tester pre-loads 1 year -> cache full from the first row)
ProcessStructure : begin = max(startBar, 2*len) = max(100, 700-600) = 100
             loop i in [100, total)  -> 600 bars ANALYZED (cache bars 0..99 = warmup)
swing pivot: valid absolute p in [r-649, r-50]   (rel [50, 649])
             (IsPivotHigh needs neighbors p±50 inside the 700-bar cache)
g_swHigh   : price of the LAST swing HIGH pivot in that window; 0 if none
g_swLow    : analogous
```

## 2. Quantitative evidence (ml/parity/verify_window.py)

```text
f6_eqpos  : mismatch = 0/14850  max|d| = 0.000000
f14_dhigh : mismatch = 0/14850
f15_dlow  : mismatch = 0/14850
f17_range : mismatch = 0/14850
```

The windowed reconstruction (sw_high/sw_low from [r-649, r-50], 0 when empty)
reproduces the EA 100% for all four window-sensitive features.

---

## 3. Minimum history requirement per feature

| Feature ID | Minimum lookback | Warm-up | State dependency | Runtime window |
|---|---|---|---|---|
| f0/f1/f2 (HTF bias) | 200 newest HTF bars | need<120 -> 0 | none (200-bar window) | HTF cache 250, corrected: 200 newest ending at E_ea |
| f3/f4/f5 (trend/bias) | 600 bars of structure analyzed | begin=100 (cache idx) | pivot feed from [r-649, r-50] | M15 cache 700 |
| f6 eq_pos | last swing pivot in [r-649, r-50] | no pivot -> 0 | sw_high/sw_low | M15 cache 700 |
| f7/f8/f9 (sweep/choch) | 600 bars of internal structure | begin=100 | swing timeline | M15 cache 700 |
| f10/f11 (EQH/EQL) | pairs in [r-649, r-50] | pair < EQ_BARS skip | swing pivot list | M15 cache 700 |
| f12/f13 (delta) | 10 bars (DELTA_BARS) | kk<=0 -> 0 | none | 10 bars |
| f14/f15 (dist) | sw_high/sw_low + ATR(14) | sw=0 -> 0 | structure | M15 cache 700 |
| f16 (mom20) | 21 bars (close[tc-21]) | tc<21 -> 0 | none | 21 bars |
| f17 (range) | sw_high/sw_low | rng=0 -> 0 | structure | M15 cache 700 |
| f18 (conf) | derived (f0-f2,f5,f6,f7,f8) | — | — | follows the above |
| ATR | 14 bars | i>=13 | none | 14 bars |

**Technical finding**: a LOOKBACK of 600 alone is NOT sufficient — pivots at rel [50,99]
(absolute [r-649, r-600]) also determine sw_high/sw_low and the EQH/EQL pairs,
and their confirmation uses cache bars 0..99 (warmup). The minimum requirement for
full parity = **700-bar cache** (600 analyzed + 100 warmup).

---

## 4. Decision

```text
Minimum required : 700 M15 bars (cache) — 600 analysis bars + 100 warmup;
                   HTF: 250-bar cache, 200 used (P2.1 corrected)
Production window: cache 700 / HTF 250 (runtime deployment window = SOURCE OF TRUTH)
Training window  : must replicate the 700-bar cache ending at the row bar (begin=100),
                   HTF 200 newest ending at E_ea — implemented in P2.5
Parity           : f6/f14/f15/f17 = 0/14850 (windowed reconstruction)
```

Note: the window was NOT chosen based on AUC. The runtime deployment window
is the contract; if training needs a different window, that is outside P2
(recorded as a future experiment).