Related: Project overview · OOS dashboard · Redaction report
Every number in the case study (README.md) traces to one of three sources in the private research repository:
Cross-check: R and F agree on every overlapping figure (per-year totals, Sharpe, drawdowns) to the fourth decimal.
| Figure | Value | Source |
|---|---|---|
| 2026 YTD total (2026-01-01 → 2026-06-10) | +45.1% | R per_year.2026.total = F metrics.total (0.4508) |
| 2026 YTD Sharpe (annualized, 4h bars, n=961) | 2.25 | R = F |
| 2026 YTD Sharpe standard error (Lo 2002) | ±1.51 → 95% CI ≈ [−0.7, +5.2] | R per_year.2026.sharpe_se |
| 2026 YTD max drawdown | −21.3% | R = F |
| 2026 BTC / ETH buy-and-hold | −29.7% / −44.9% | R bench.2026 = F bench |
| 2025 frozen replay | +54.5%, Sharpe 1.52, maxDD −11.9% | R per_year.2025 = F sanity |
| 2025 BTC / ETH | −6.6% / −11.4% | R bench.2025 |
| 2024 frozen replay | +16.5%, Sharpe 0.60, maxDD −28.0% | R per_year.2024 = F sanity |
| 2024 BTC / ETH | +121.1% / +46.8% | R bench.2024 (system underperformed BTC in 2024 — stated in README) |
| 2026 per-book: BTC +54.0% (Sh 2.39), ETH +35.3% (Sh 1.82) | F books |
| Figure | Value | Source |
|---|---|---|
| Total return | +144.5% | R full_period.total |
| CAGR (2.44 years) | +44.2%/yr | R full_period.cagr |
| Sharpe (n=5,347 bars) | 1.20 ± 0.64 (SE) | R full_period.sharpe, .sharpe_se |
| Max drawdown | −28.0% | R full_period.max_dd |
| Realized vol | 35.7%/yr | R full_period.vol |
| BTC / ETH same window | +46.0% / −27.8% | R full_period.bench_total |
Note: per-year numbers are standalone replays (selection restarts each Jan 1); the full-period run is continuous. January boundary weeks can differ slightly between the two; both are reported as what they are.
| Figure | Value | Source |
|---|---|---|
| 2026 / 2025 / 2024 unleveraged | +30.2% / +37.4% / +12.6% | R reference_unlevered |
| Drawdowns unleveraged | −15.2% / −8.4% / −20.3% | R reference_unlevered |
The shipped configuration sizes ~1.4× the unleveraged variant at the same Sharpe (2.25/1.52/0.60 vs 2.22/1.54/0.59). The dial is a risk choice.
All from R graveyard.* (per-year totals + delta_vs_base); chart shows
the worst single-year delta.
| Rejected idea | Worst-year delta | All-year deltas (2024/2025/2026) |
|---|---|---|
| Daily refit cadence | −40.9pp (2025) | −33.7 / −40.9 / −25.9 |
| Daily bars | −45.7pp (2025) | −11.8 / −45.7 / −14.8 |
| 1h bars | −8.6pp (2024) | −8.6 / +6.5 / −1.4 (mixed — see discrepancy log) |
| Drawdown kill switch (12%) | −43.1pp (2026) | −5.6 / −17.8 / −43.1 |
| Emergency refit on 5% drawdown | −22.1pp (2026) | −3.4 / −2.4 / −22.1 |
| Sentiment + positioning signals in the zoo | −22.8pp (2025) | −13.6 / −22.8 / +6.6 (fails 2 of 3 years) |
| Fixed leverage ≥1.5x without vol scaling | account → $0 | B graveyard: five configs, each with a triple-digit best year, all liquidated (four on 2020-03-12, one on 2018-01-16) |
Supporting artifact figures quoted in README:
| Figure | Value | Source |
|---|---|---|
| Kill switch on/off (top-1 unleveraged variant) | 2025: +41.0% → −12.6%; 2024: −6.3% → −20.4% | F, kill-switch test artifact |
| Meta-grid honesty | 108 cells, 25% positive in 2026, median −17.9%, worst −41.5% | F meta |
| Weekly hit rate / avg win / avg loss | 52% / +4.1% / −2.7% (126 weeks 2024→now, measured on the unleveraged-era config) | F notes (quoted); recomputation procedure described there |
| Refit-weekday dispersion | up to ~33pp/yr; Monday top-2 in all 3 years | F notes (quoted from the timing-luck study) |
| Cross-venue agreement | Binance vs OKX median 0.6bp, vs Bybit ~4.5bp; results identical to the cent | F notes (quoted); reproducible via the cross-venue fetch + blend test |
| Figure | Value | Source |
|---|---|---|
| Cost per side | 0.15% on every position change | fit pipeline COST constant; charged on the net position delta (switches between similar configs net correctly — verified in the position-stitching step) |
| Funding | real timestamped Binance events, signed | funding loader; interval-agnostic by construction (events summed into containing bar). BTC/ETH stayed at 3 events/day through the 2025-09-18 formula change — verified by event counts in the funding data set |
| Binance VIP-0 fees (for the “0.15% is conservative” claim) | spot 0.1%, USDT-perp taker 0.05% | external research (mid-2026), from Binance’s published VIP-0 fee schedule |
This log reconciles figures from earlier development notes against the numbers the current code actually produces. It exists so that any older claim a reader might encounter can be traced to why the shipped figure differs. The retired early figures themselves are not published anywhere in the case study; only the current, reproducible numbers are.