Crypto Arbitrage Bot

I wanted to know whether the arbitrage opportunities crypto sites advertise are real once fees, spreads and slippage are counted. So I built the system to find out: a trading platform that watches the whole market around the clock and trades it on paper money, with every cost modelled.

• Funding-rate dashboards show headline yields of hundreds of percent a year. Are any of them real?
• Do the same coins trade at different prices on different exchanges for long enough to capture?
• Can the same sports game be priced differently on two prediction markets?
• Can a system learn from its own trades and get better over time?

• Live market data straight from 16 exchanges' own public APIs (centralised and decentralised): bid, ask, funding, volume and open interest, refreshed every few seconds
• Funding carry: long spot, short perpetual, collecting funding, backtested on the full funding history since 2023
• Cross-exchange funding arbitrage: short the exchange paying more, long the one paying less
• Price-gap arbitrage: buy the cheaper venue, sell the dearer one, only when the gap beats every fee and spread
• Prediction-market arbitrage: the same game matched across two venues, traded only when both sides together cost less than the payout
• A self-learning layer that records the conditions of every trade and trains a model on the results, running in watch-only mode until it beats the rules

• Every trade pays modelled taker fees, bid-ask spread and slippage on both legs
• Funding is only credited from each exchange's own settlement history, never guessed after an outage
• A kill switch, per-coin cooldowns after losses, risk limits and an audit log on every admin action
• Live trading is refused in code: going live is a manual decision, made only after out-of-sample proof

Most advertised "opportunities" disappear once real costs are counted. One spread looked like 700% a year, but the two instruments weren't actually the same. The fast funding strategy lost money to costs and was redesigned twice. Price gaps between 15 and 60 basis points were the only band that paid, and slow funding carry earns a small, steady return. Finding that out on paper costs nothing; finding it out with real money doesn't.

The same discipline runs my ad accounts: measure every cost, trust only results you can verify, and don't scale anything until the evidence says it works.