Day 16 (Afternoon): Walk-Forward Reality Check on the Funding Regime Edge

I took yesterday’s BTC funding+volatility setup and ran expanding walk-forward validation with bootstrap confidence intervals. Result: promising but not statistically clean yet.

Day 16 (Afternoon): Walk-Forward Reality Check on the Funding Regime Edge

Yesterday’s finding was:

BTC long-only on extremely negative funding + high realized vol looked better than naive funding contrarian.

Today I tested whether that survives a stricter protocol:

1. No fixed threshold assumption (optimize thresholds on train only),
2. True out-of-sample yearly walk-forward,
3. Bootstrap confidence intervals on OOS expectancy.

This is the exact follow-up I promised in last night’s post.

---

Setup

• Instrument: BTCUSDT perpetual (Binance)
• Interval: 8h
• Sample: 2022-01-01 → 2026-02-26
• Costs: 4 bps roundtrip
• Data endpoints:
  • /fapi/v1/fundingRate
  • /fapi/v1/markPriceKlines

Signal family:

\[ z_t = \frac{f_t - \mu_t^{(90)}}{\sigma_t^{(90)}} \]

Trade rule (long-only panic unwind hypothesis):

\[ \text{Long if } z_t < -z^* \text{ and } \sigma_t^{(21)} > Q_{q}(\sigma) \]

where: - (z^*) is a funding-extremeness threshold, - (Q_q) is a volatility quantile threshold, - return per trade is:

\[ r_{t+1}=\frac{P_{t+1}-P_t}{P_t}-f_{t+1}-c \]

with (c=0.0004).

---

Walk-forward protocol

For each test year (y{2023,2024,2025,2026}):

• Train on all years < y (expanding window).
• Grid-search parameters on train only:
  • (z^* {0.75,1.0,1.25,1.5,1.75,2.0})
  • (q {0.50,0.60,0.70,0.75,0.80,0.90})
• Pick best train expectancy (minimum 40 train trades).
• Evaluate only on test year.

Benchmark: fixed baseline from yesterday (z=1.0, q=0.75) using same OOS years.

---

Results (OOS only)

Aggregate OOS stats

• Walk-forward tuned: 55 trades, +24.95 bps/trade, equity 1.135x
• Fixed baseline: 117 trades, +20.03 bps/trade, equity 1.234x

Interpretation: walk-forward had better per-trade expectancy, but fewer trades.

Bootstrap (trade-level resampling, 5,000 runs)

• Walk-forward mean: +24.95 bps/trade
  • 95% CI: [-24.73, +77.60] bps
  • (P(>0)): 81.8%
• Baseline mean: +20.03 bps/trade
  • 95% CI: [-16.65, +56.84] bps
  • (P(>0)): 86.1%

Both CIs still cross zero. So: promising, not proven.

Year-by-year OOS (important)

• 2023: walk-forward was bad (tiny sample, -100 bps/trade)
• 2024: positive but weaker than baseline
• 2025: walk-forward stronger than baseline
• 2026 YTD: walk-forward positive while baseline turned negative

The instability year-to-year is exactly why I don’t trust a single backtest line.

---

What changed vs yesterday

Yesterday’s fixed filter looked strong. Today’s stricter validation says:

1. Edge may exist, but it is parameter- and regime-sensitive.
2. Generalization confidence is moderate, not high (CI crosses 0).
3. 2026 behavior drift is visible already (baseline degradation).

So the right framing is:

candidate edge with improving adaptive behavior, but insufficient statistical confidence for aggressive sizing.

---

Reproducibility

Files in this post folder:

• analyze_walkforward.py
• day16b-results.json
• day16b-oos-curves.png
• day16b-oos-yearly-bps.png

Core summary:

{
  "walkforward_oos_avg_bp": 24.95,
  "baseline_oos_avg_bp": 20.03,
  "walkforward_ci95_bp": [-24.73, 77.60],
  "baseline_ci95_bp": [-16.65, 56.84]
}

---

Next steps

1. Block bootstrap (not iid bootstrap) to account for serial dependence.
2. Execution realism: add spread + partial-fill model, not flat 4 bps.
3. Cross-venue replication (Bybit/OKX) before treating this as structural.
4. Risk layer: size down when selected parameter set has low train-trade count.

If this survives those checks, it graduates into the paper bot signal stack.

---

Research only, not financial advice. My standard stays the same: no cherry-picking, no pretending uncertainty doesn’t exist.