milliways01
The production system at the centre of the research programme: a market-neutral, cross-sectional portfolio on Binance perpetuals. This page covers the logic of the system at a conceptual level, the infrastructure of the live engine, and the historical simulation results.
How it works
Every day, roughly the hundred most liquid perpetuals are selected using only the liquidity data known that day. Two families of signals rank these assets against each other: one comes from order flow, meaning how aggressive buyers and sellers are; the other comes from the funding market, meaning how crowded leveraged positioning has become.
The portfolio takes positions at both ends and is kept neutral to the market's direction. Rebalancing isn't tied to a single calendar: it averages sub-portfolios that start on different days, so the result doesn't rest on a lucky choice of schedule.
Two signal families
Order flow and funding: two different sources of information with almost no correlation between them.
Market-neutral construction
Dollar-balanced and beta-neutral, with a cap on how much any one asset can take.
Phase-averaged rebalancing
An average of sub-portfolios that start on different days, so there is no calendar luck.
The live engine and data infrastructure
The system is made of processes that run independently and talk to each other only through files, so if one fails the others keep going. Data that can't be downloaded later (order book snapshots, liquidations) is recorded in real time, and every night the raw data is verified, reconciled against the archive and backed up.
- 01
Market data
Candles, funding, mark price, order book depth, liquidations.
- 02
Recorder
Records data that can't be downloaded later, in real time.
- 03
Decision engine
Daily cross-sectional decision: target weights and risk limits.
- 04
Execution model
Fills are modelled against the depth of the live order book.
- 05
Data validation and archive
Each day's data is checked against the exchange's official archive, sealed with a tamper-evident SHA256 fingerprint and backed up. How closely the engine's decisions track the simulation is reported as well.
- 06
Monitoring
Read-only dashboard: positions, ledgers, system health.
Safety layers
- Separation of permissionsEach component has only the access it needs.
- Static code scanningThe code is scanned for disallowed calls before the system starts.
- Pre-start auditThe engine won't run unless the safety audit comes back clean.
Historical simulation, 2020–2025
The simulation uses unit-based positions, measured spreads and order book impact, time-stamped funding payments, and includes assets that were later delisted. Every figure is after costs.
Historical simulation, 2020–2025 · model v1 · realistic cost and execution model
- Sharpe ratio (95% confidence interval 0.93–2.59)
- 1.76
- annual compound return
- 25.3%
- maximum drawdown
- −14.4%
- annual volatility
- 13.3%
What happens if costs rise
Sharpe ratio by cost multiplier (post-warm-up period). The default cost is about 6 basis points per side; breakeven is roughly 11 times that, 66 basis points per side. The 1× point is 1.81; the headline 1.76 includes the warm-up period.