Since 1 January 2026 I have been running an AI-driven crypto trading strategy live on Binance. This post explains how it is built and where you can inspect the results yourself.
You can see the strategy and its live statistics at aicrypto42.io, with the full performance dashboard at aicrypto42.io/entries.
What problem does an AI trading model actually solve?
Most retail trading strategies are a stack of hard-coded rules. If RSI is below 30 and price crosses the 50 EMA and volume is above average, then buy.
The problem is not that those rules are wrong. It is that they are fixed weights chosen by a human, applied identically to every market condition. RSI below 30 means something different in a trending market than in a ranging one, and something different again for BTC than for PEPE.
A model solves a narrow version of this: instead of you deciding how much each indicator matters, the model learns it from what actually happened.
How does the model work?
The core is an XGBoost classifier. Dozens of technical indicators and engineered features go in. A single probability comes out.
That probability replaces the entire if-then tree. Rather than asking “did all six of my conditions fire?”, the strategy asks “given everything I can currently measure, how likely is this setup to resolve profitably?”
This matters for a practical reason. Hard-coded rules combine badly — every condition you add makes the system fire less often, until it trades once a month. A model combines signals multiplicatively rather than restrictively: weak evidence from six sources can add up to a confident prediction, which a rule-based system would have thrown away.
Why require a second confirmation?
Because a model that is confidently wrong is worse than no model.
The strategy requires two independent confirmations before it takes a trade:
- The AI probability from the XGBoost model
- A classical trendline or breakout signal — widely used, well understood, and importantly not derived from the model
These are computed on separate timeframes: an analysis timeframe for the decision, and an entry timeframe for execution. Agreement across both a different method and a different timescale filters out a large share of the model’s false positives.
This is a deliberate design choice with a cost. The strategy trades less often than the model alone would. That is the point.
Why analyse on one timeframe and enter on another?
Because the two jobs have different requirements.
Deciding whether a setup is worth taking benefits from a slower timeframe — more context, less noise, fewer false signals. Entering benefits from the fastest timeframe available, because every tick of slippage between the decision and the fill is a direct cost.
So analysis runs on the higher timeframe and execution happens on the 1-minute interval. The strategy knows what it wants before it needs to act quickly, which is the only way to act quickly without acting carelessly.
How is risk handled?
Trailing stops and position-level risk limits are built into the strategy, not bolted on.
This is where most backtested strategies fail in production. A backtest reports the average outcome; a live account experiences the sequence. A strategy with a good average and no drawdown control will, eventually, hit the run of losses that ends it.
The dashboard therefore tracks maximum drawdown and profit factor as first-class metrics, alongside the equity curve.
What do the live numbers show?
They show whatever they show — that is the entire point of publishing them.
The dashboard at aicrypto42.io/entries reports, continuously and without curation:
- Win rate across all symbols, over 1-day, 7-day, 30-day and 1-year windows
- Equity curve, daily
- Net PnL
- Maximum drawdown and profit factor
- P&L broken down by symbol — BTC, ETH, SOL, XRP, ADA, DOGE, PEPE and TRUMP
Including the losing periods. A strategy page that only shows winning windows is marketing, and you should read it as such — mine included, which is why the numbers are live rather than screenshotted.
If you are evaluating any algorithmic strategy, the two numbers to check first are profit factor (total profit divided by total loss — above 1.0 is profitable) and maximum drawdown (the worst peak-to-trough fall). A high win rate with a profit factor below 1.0 means many small wins and a few catastrophic losses.
What I would tell someone building this
Three things, in order of how much time they will save you:
- Feature engineering beats model selection. The gap between a mediocre and a good XGBoost model here was almost entirely in how features were constructed, not in hyperparameters.
- Backtest realistically or not at all. Fees, slippage, and the fact that your fill is not the price you saw. Optimistic backtests are the single most common reason live results diverge from expectations.
- Publish the results. It forces honesty about what the system actually does, and it is the only evidence anyone should accept.
This post describes an engineering project. It is not financial advice, and nothing here is a recommendation to trade or invest. Algorithmic trading carries substantial risk of loss, and past performance does not indicate future results.
See it live: aicrypto42.io · Live statistics: aicrypto42.io/entries
