BOTLAB in 5 minutes

An AI-run quant lab that builds crypto trading strategies, tests them against real costs and strict statistics, and publicly retires the ones that fail, including its own.

Try the free strategy checker See every bot's results Watch the 3-minute video

The problem

Most trading bots sold online show backtests that ignore fees and slippage, peek at future candles, or were picked as the best of hundreds of tries. They look great and then lose money. Buyers have no simple way to tell.

What BOTLAB does

1. AI agents propose and build strategiesAn idea goes through an AI critic, then a backtest with real exchange costs (taker fees, order-book slippage, funding).

2. Every result faces the same six gatesSample size, probabilistic and deflated Sharpe ratio (adjusted for how many variants were tried), recent performance, concentration (no profit from just 2 days or 3 coins), side leak and drawdown. Look-ahead is audited by recomputing indicators bar by bar.

3. Survivors go to paper trading, failures go to a public graveyardPaper books keep running on live data. Anything that fails is listed with the reason, so nobody repeats it.

4. Real money only behind guard rails, with a human owner's approvalTwo bots trade real money today: combo LIVE (a Bybit sub-account that copies the best paper books) and paper1m LIVE (the main account since 23 Oct 2026, after the original Atlas bot was retired). Both results, losses included, are on the Fleet page. New filters start as shadow tests, and AI proposals only go live after the owner approves them.

5. The whole fleet runs itselfScheduled agents publish results, a monitor checks every bot every 5 minutes and restarts silent ones, and code updates reach the server only when signed.

Numbers as of 26 Oct 2026

204

strategy books tracked on the site, each through the six gates

book(s) pass all six gates and are marked ready for micro-live: mr4h/long (110 paper trades). 202 are not ready. We say so.

8,935

strategy variants tested, all counted in the multiple-testing penalty

exchange-tool tests (funding, open interest, long/short ratio, basis) across 15m/30m/4h/1D. Only one tool held up out of sample (study, 11 Oct 2026).

look-ahead mismatches when 48 random trades × 12 tools were recomputed from the live API (study, 11 Oct 2026)

All performance figures on BOTLAB are paper or backtest results unless marked as real money. Past results do not predict future returns. Nothing here is investment advice.

Try it yourself (2 minutes)

What	How
Check a strategy	Open the checker, press Use sample data or paste a CSV with one result per trade. You get PSR, DSR, drawdown and the six gates, computed in your browser. Use Copy share link to share only the summary.
See honest failures	Results lists every book, and the Bot graveyard tab shows retired strategies and why.
Read the research	In Results, open the research panels: trend tools on Atlas (real money), tools on the 4h bot, and exchange data tools on four timeframes.
Run the gates yourself	The six gates, the cost model, PSR / Deflated Sharpe / MinTRL and the look-ahead audit are MIT-licensed Python at github.com/botlabtrade/botlab-oss. Install it, point it at your own trades, and check every number on this site against your own run.
Paper trade	Paper trading with 10,000 virtual USDT. Sign in with Google.

Architecture

Bybit public market data ──► research & paper bots (VPS, ~110 scheduled tasks)
                               │  real-cost model · 6 gates · look-ahead audit · DSR
                               ▼
                  strategy ranking + live gate ──► signed sync ──► Cloudflare Worker + D1
                               │                                     │
             5-minute monitor (bot_pulse) + watchdog          botlab.org.uk (static site,
             Telegram alerts, auto-restart                     checker, paper trading, accounts)
                               │
             real-money bots (owner-approved): combo LIVE (sub-account) · paper1m LIVE (main account)

What we are honest about

No strategy has yet passed every gate on a large sample. The site shows this on the first screen.
15-minute and 30-minute signals lose money after costs in our tests, in every month we checked. We publish that too.
The checker cannot see your code, so it cannot detect look-ahead in it. It says so in every report.

Contact: through the chat on botlab.org.uk or contact@botlab.org.uk.

Home · Platform · Fleet · Trading · Strategy audit · Learn · Lab News · Privacy