edge-optimize

name: edge-optimize description: Full 4-step systematic optimization of 8GSR gate thresholds per docs/8GSR_TASK.md. Runs univariate scan → multivariate grid search → time-based cross-validation → ranked top-10 parameter sets. Use when the user says "optimize the strategy", "find the best parameters", "run the full optimization", or "8GSR task".

Edge Optimize — Full 8GSR Parameter Optimization

You are running the full systematic optimization described in docs/8GSR_TASK.md. Your goal: find the parameter combination that produces the highest statistically significant, profitable edge at 120s expiry.

Step 1: Health Check

Call po_health. If bot DB not accessible → STOP.

Step 2: Univariate Scan

Call po_find_edge.

This returns all 13+ dimensions bucketed with full statistics including z-score, p-value, and Wilson 95% CI (now that bStats is upgraded).

For each dimension, extract:

• The best bucket per parameter per direction (from best_thresholds)
• Any bucket with p_value < 0.10 — note as statistically significant
• Any bucket with p_value < 0.05 — note as strong edge

Document top 2-3 ranges per parameter:

Focus on 120s expiry (exp_120s). Flag any dimension where best bucket WR < 50% — those parameters add no edge.

Step 3: Multivariate Grid Search

Call po_grid_search (direction='both').

This tests ALL combinations of:

• stc_prev (6 thresholds per direction)
• stc_delta (5 thresholds per direction)
• g3_depth (5 thresholds per direction)
• g1_bars_ago (1, 2, 3, any)

= up to 600 CALL + 720 PUT combinations, filtered to n≥20.

From the results:

1. List top 10 combinations per direction (by win rate)
2. Note the p_value and Wilson CI for each — only trust combinations with p_value < 0.10
3. Check if the current live gate params appear in the top results — confirm they are optimal or find a better set

Step 4: Cross-Validation

From po_find_edge, the cross_validation section splits the 120s signal pool at the median timestamp into in-sample (first half) and out-of-sample (second half).

Check:

• Does the overall WR hold within 5% between halves?
• Does the best parameter set from Step 3 perform consistently across both halves?
• If out-of-sample WR drops >10% vs in-sample → mark as OVERFIT RISK

To validate a specific parameter set across time: Call po_simulate with those params to see overall performance, then manually compare po_find_edge cross_validation in_sample vs out_of_sample WR for that direction.

Step 5: Rank and Report

Build a top-10 table from the grid search results. For each entry:

Select the single best set using this criteria (in order):

1. p_value < 0.05 (statistically significant)
2. n ≥ 30 (enough trades)
3. Profit factor > 1.2
4. Wilson CI lower bound > 50% (confidently above breakeven)
5. If tied: prefer the set with higher n

Step 6: Compare vs Current Live Gates

Current live gates in bot/indicators.js:

• CALL: stcPrev ≤ 25, delta 0–0.5, g3_depth < -150
• PUT: stcPrev ≥ 85, delta -0.9–0, g3_depth > 175

For each direction, state:

• Are the live gates already at the optimal grid point?
• If not, what change would improve WR and by how much?
• Quantify the improvement using the grid search results

Step 7: Simulate Recommended Changes

For any gate change identified, call po_simulate with those params to confirm the improvement vs baseline before recommending a bot update.

Step 8: Write Report

Write docs/edge_report_[YYYY-MM-DD].md with:

# 8GSR Optimization Report — [date]

## Dataset
- Total loose-pool signals: N
- Validated at 120s: N
- Date range: [first signal ts] to [last signal ts]

## Univariate Findings
[Table of best bucket per parameter — WR, n, z_score, p_value]

## Top 10 Parameter Sets — CALL
[Table from grid search]

## Top 10 Parameter Sets — PUT
[Table from grid search]

## Cross-Validation
- In-sample WR:     N% (n=N)
- Out-of-sample WR: N% (n=N)
- Verdict: ROBUST / MARGINAL / OVERFIT RISK

## Recommended Best Set
**CALL:** stc_prev ≤ N, delta < N, g3_depth < N, g1_bars_ago = N
  WR = N%, n = N, z = N, p = N, Wilson CI = [N, N]

**PUT:** stc_prev ≥ N, delta ≥ N, g3_depth > N, g1_bars_ago = N
  WR = N%, n = N, z = N, p = N, Wilson CI = [N, N]

## Gate Changes Required
| Param | Current | Recommended | Delta WR | Delta PnL |
|---|---|---|---|---|

## Next Steps
1. Apply recommended gate changes via po_simulate first
2. If confirmed: update bot/indicators.js
3. Accumulate N live trades then re-run /edge-optimize to validate

Rules

• Never recommend a gate change without po_simulate confirmation
• Flag any finding with n < 20 as "insufficient sample"
• p_value < 0.10 = MARGINAL, < 0.05 = SIGNIFICANT, < 0.01 = STRONG
• Wilson CI lower bound > 0.50 = confidently above breakeven
• Always separate CALL and PUT — different optimal thresholds are expected
• If no combination passes p < 0.10 with n ≥ 20, report "no statistically significant edge found" and recommend accumulating more data