All work

BTC Strategy Lab

A research bench whose default answer is 'not yet'.

Context
Personal project
Year
2026
Role
Sole author
Stack
Rust, tokio, axum, SQLite

The problem

Most backtests are overfit by the person running them. I wanted a lab that tells me honestly which strategies still make money after real costs, across BTC, NIFTY, BANK NIFTY and NSE stocks, without ever placing a real order.

  • Out-of-sample windows are sealed and read once.
  • Fills come from recorded order books, never assumed.
  • Rules for retiring a strategy are written before the evidence arrives.

Architecture

Fig. 1 A recorder stores the order book every second, which drives the fill model. Walk-forward research runs on sealed out-of-sample windows, and only strategies that clear a deflated Sharpe gate of 0.95 become paper books. Losing books are retired by rules written in advance.

What I built

  1. 01A 10-crate Rust workspace: data, India markets, stocks, strategies, simulator, research, live, experiments and the app.
  2. 02A live recorder that stores the L2 order book every second and the option chain every minute as compressed JSONL.
  3. 03A fill model driven by those recordings; futures without broker depth are tagged rather than faked.
  4. 04Proof gates in order: walk-forward, a parameter-neighbour check, a 10,000-resample block bootstrap, a deflated Sharpe probability of at least 0.95 and sealed one-shot holdouts.
  5. 05Every registered strategy runs from day one as a practice paper book on live data; only books that clear the deflated Sharpe gate may draw on the shared paper account. A React instrument panel watches them.
BTC Strategy Lab's Today view: the shared paper pool's balance, profit and loss since start, free margin, open risk to stops, and a table of open positions with stops and holding times.Paper only · no real orders
Paper books before the 16 Nov 2026 retirement
The Experiments page: plain-language research findings on exit efficiency and bar-start sensitivity, above a table of trend books with trade counts and how much of each winner's best move was kept.Paper only · no real orders

By the numbers

logged trials, four markets
5,499
losing books retired
59
Rust tests
2,729
deflated Sharpe probability to pass
0.95

Hard problems

01

1.67 looked great until it met 6.28

The best trusted out-of-sample BTC Sharpe was 1.67, over 790 days. But 1,252 BTC trials had been logged, and given how widely their Sharpes spread, the best of that many would reach about 6.3 by luck alone. Its deflated Sharpe was 0.000. That gap is why the gate exists, and why the playground on this site counts your attempts.

02

Killing my own strategies on schedule

On 16 Nov 2026, 57 of 117 paper books and 2 experiment books were retired under a rule written in advance: sealed out-of-sample expectancy at or below zero after costs. About 40 strategies have now been through the full proof pipeline and none has passed after costs on unseen data. Every main paper account sits in cash, which is the correct answer when nothing has earned the right to trade.

03

Four markets, fourteen phases

It grew from BTC-only to NIFTY, BANK NIFTY and NSE stocks with per-market paper accounts and date-correct Indian costs, across 14 research phases. Along the way the pipeline caught a position-sizing look-ahead and a futures-roll bug before any sealed data was read.