Machine learning (ML) shows up throughout crypto trading in several distinct layers, from the infrastructure that executes trades to the models that try to predict markets. Here's a practical breakdown:
Price prediction / signal generation
- Models feed on historical price, order-book depth, on-chain data (whale movements, exchange inflows), and sentiment (news, social media) to output a directional signal — up/down, or a probability of a move.
- Common approaches: time-series models (LSTM/GRU networks, transformers), gradient-boosted trees (XGBoost, LightGBM) for tabular feature sets, and reinforcement learning for agents that learn trading policies from reward signals.
Execution optimization
- This is arguably the most reliable real-world use. ML helps break up large orders to minimize market impact — deciding when and how much to trade at each tick, not just what direction. Market-making bots use it to manage inventory risk and quote pricing.
Risk management
- Position sizing, volatility forecasting (e.g. using ML to estimate realized vol), anomaly detection to flag unusual moves or flash-crash risk, and portfolio rebalancing based on predicted correlations.
On-chain and alternative data
- Classifying wallet behavior, detecting accumulation/distribution patterns, or clustering addresses into "smart money" vs. retail. NLP models read news headlines and social sentiment as a feature.
The honest caveats
- Markets are noisy and non-stationary — a model that works for months can decay as regime shifts. Overfitting to backtests is the classic failure mode.
- Many "ML hedge funds" and crypto trading bots market themselves heavily; the edge usually comes from the data pipeline and execution, not the model architecture itself.
- Fees, slippage, and latency often eat the theoretical edge. There's no reliable free data source showing which ML strategies actually profit.
This isn't financial advice — it's a description of how the techniques are applied, not a signal on what to trade.
If you're curious about the practical side, I can look up where the current data sits — e.g. the trending coins list for today, or check a specific project's live market data. Want me to pull something specific?