Dark Matter

Pulsar

ML-driven statistical arbitrage · Crypto perpetuals

BacktestBacktest · Binance · 730 days · net of costs · in-sample and out-of-sample30 Sep 2024 – 30 Sep 2026

Equity · cumulative % from the start of the record

0%+10%+20%+30%Oct ’242025AprJulOct2026AprJulOut-of-sample: held out of the rule search

Shaded: out of sample, days never used to build or tune the strategy. Band: the middle 90% of 2,000 paths with the same daily returns in a resampled order.

Statistics

Total return
+27.7%
CAGR
+13.0%
Sharpe
2.15annualised
Sortino
3.21annualised
Max drawdown
−4.15%peak to trough, % of starting capital
Worst day
−3.52%9 May 2025
Worst month
−1.34%Feb 2026
Days
730

Total return, Sharpe, max drawdown and days are the record’s own figures. Sortino, worst day, worst month, CAGR are computed on this page from its daily series, annualised at the series’ own 365 observations a year.

In sample and out of sample

StretchDatesReturnSharpeMax drawdownDays
In-sample: the months the rule was chosen on30 Sep 2024 – 31 Mar 2026+22.1%2.07−4.1%547
Out-of-sample: held out of the rule search1 Apr 2026 – 30 Sep 2026+5.7%3.03−1.1%183

Monte Carlo · 2,000 resampled paths

Over the nextChance of a lossReturn, 5th / 50th / 95th percentileMax drawdown, typical / 1 in 20
90 days4%+0.1% / +2.6% / +6.0%−1.1% / −1.9%
180 days<1%+1.7% / +5.5% / +9.8%−1.2% / −2.1%
365 days<1%+5.7% / +11.0% / +17.3%−1.5% / −2.4%

Daily returns of the out-of-sample days (183 days), resampled in blocks of about ten days. Resampling reorders the days this record had; it cannot add a market it never saw.

Drawdown · peak to trough, % of starting capital

−4%−2%0%Oct ’242025AprJulOct2026AprJul

Rolling Sharpe · 63-day window, annualised

052025AprJulOct2026AprJul

Monthly returns

YearJanFebMarAprMayJunJulAugSepOctNovDec
2024no datano datano datano datano datano datano datano data0.0%+2.3%+3.0%+3.3%
2025−0.3%+1.2%+1.3%+1.0%−0.8%+1.0%+2.1%+2.4%+0.9%+0.1%+2.8%+1.8%
2026+1.1%−1.3%+0.1%−0.2%0.0%+3.4%+0.2%+2.1%+0.1%no datano datano data

Outlined months are partial: the record starts or ends inside them.

How this was tested

  • This is a backtest: the strategy simulated over historical market data. No money was traded.
  • It covers 30 Sep 2024 to 30 Sep 2026, 730 days in the record.
  • Returns are measured as % of starting capital.
  • Costs: 5-minute checks at the mid, 10-minute TWAP fills. Half the fills at the maker fee, half at the taker fee plus 2 bps slippage; funding charged on both legs. $8,000 book: 20 pairs at $200 a leg. No compounding. Open trades are marked to market every hour.
  • Partly out-of-sample: the shaded stretches, 183 of the 730 days, were never used to build or tune the strategy; the rest overlaps the data it was tuned on, which tends to flatter the result.

Record: Backtest · Binance · 730 days · net of costs · in-sample and out-of-sample. Exported 14 Oct 2026 from platform commit c6e04beb (ML-Stat-Arb-Signal-Service ops/pulsar_intrabar.py (branch pulsar-intrabar-5m): monthly point-in-time screen, delisted coins included, 2543 trades).