Every millisecond counts. In HFT market making, a 1 ms latency decides everything — on Binance, every millisecond spread on the BTC/USDT pair with a volume of 1000 BTC yields 50–100 USD. But speed without intelligence leads to inventory accumulation on directional moves: without ML, you can lose 2000 USD per hour due to adverse selection. Our RL agents reduce the adverse selection ratio by 12–18% and increase P&L by 15–30% compared to classic static models (Avellaneda-Stoikov model).
The Economics of Market Making
P&L consists of four components: spread capture, inventory risk, adverse selection, and maker fees. The goal is to maximize spread capture while minimizing inventory risk. The classical Avellaneda-Stoikov model provides a baseline formula, but we replace fixed parameters with dynamically estimated ML.
How ML Models Reduce Adverse Selection Risk
The classical approach uses constant risk aversion and volatility. For an ML solution, we train a neural network to predict:
- Order flow intensity: how many orders will arrive in the next second.
- Adverse selection probability: the likelihood that the next order is informed.
- Short-term price direction: the price trend over the next few seconds.
Hawkes processes model the clustered arrival of orders: burst activity → temporary position reduction. In practice, this reduces the adverse selection ratio by 12–18% compared to constant reserves.
Why an RL Agent Is More Effective Than Static Rules for Inventory Management
We use an RL agent with a state: inventory, imbalance, volatility, time of day, trend. The agent selects the spread width and skew. The state space includes 6 features, and the reward function accounts for both instant profit and penalties for deviation from zero position.
# Simplified market making state space state = np.array([ inventory / max_inventory, # Normalized inventory bid_ask_imbalance, # Order book imbalance recent_vol_normalized, # Short-term volatility time_of_day_features, # Session effects recent_price_trend # Directional signal ]) action = agent.act(state) # action: [spread_half_width, inventory_skew] Stochastic control keeps inventory within given limits, asymmetrically shifting quotes to smoothly reduce the position without aggressive orders. Our engineers have 5+ years of experience in HFT and ML, ensuring stable strategy operation.
Comparison: Classic vs ML
| Parameter | Classic MM | ML MM |
|---|---|---|
| Risk aversion | Constant | Adaptive |
| Volatility | Fixed | Predicted |
| Spread | Symmetric | Asymmetric |
| Inventory control | Hard limit | RL-optimized |
| Profit | Baseline | +15–30% |
DeFi and AMM: Concentrated Liquidity
For Uniswap V3, we optimize the concentrated liquidity range. A narrower range = more fees, but higher impermanent loss risk. The RL agent decides when to rebalance the position by predicting the expected IL over a given horizon. In our tests on ETH/USDC 0.3%, IL prediction accuracy reached 89%. For an accurate assessment of your strategy and volumes, contact us — we will prepare a demo account.
Common mistakes when implementing an MM bot
- Ignoring the maker-taker fee structure: on some exchanges (Binance, Bybit), the maker rebate may exceed the taker fee, changing the economics.
- Using the same spread for all trading pairs — different liquidity requires different settings.
- Lack of latency monitoring: a delay >10 ms on a CEX destroys the ML advantage.
Comparison of CEX and DEX Market Making
| Parameter | CEX | DEX |
|---|---|---|
| Latency | <1 ms (colocation) | 1-5 blocks (15-30 s) |
| Fee model | Maker rebate | Pool fee (0.01-1%) |
| Risks | Exchange failure | Smart contract, MEV |
| Inventory | Direct | Via AMM (automatic) |
| ML focus | Order flow | IL prediction |
Technical Infrastructure
Latency is a key factor. On CEX we use colocation, on DEX — MEV-protected RPC (Flashbots). Stack: WebSocket → Order Book → Quote Engine → ML inference → Order Manager → Exchange API. We maintain inference latency under <5 ms thanks to ONNX Runtime and GPU inference on T4. Our system processes up to 10,000 orders per minute and supports 50+ trading pairs. Monitoring: P&L attribution, fill rate, adverse selection ratio, inventory over time. All metrics are streamed into Grafana with alerts on Sharpe ratio and VaR.
How We Develop an AI Bot: Step-by-Step Process
- Analysis of trading pairs and liquidity — assessment of spreads, volumes, and competitive environment.
- Selection and configuration of ML architecture — PyTorch, Hugging Face Transformers, optionally LangChain.
- Model training on historical data — backtesting with metrics (Sharpe, Sortino, maximum drawdown).
- Exchange integration — REST/WebSocket API, WebSocket order book, reconnection mechanism.
- Server deployment — Docker container, Prometheus+Grafana monitoring, alerts.
- Documentation and team training — code handover, model card, runbooks.
- 3-month warranty support — bug fixes, retraining when market conditions change.
What Is Included in the Work (Deliverables)
- Source code of the ML model with comments and configuration.
- API integration with selected exchanges (CEX/DEX).
- Server deployment with automatic restart and monitoring.
- Documentation: architecture description, launch instructions, operator runbook.
- Client team training (2–5 hours of onboarding).
- 3-month warranty support: bug fixes, model retraining when market conditions change.
Timelines: from 4 to 8 weeks depending on complexity. For an accurate estimate, contact us — we will analyze your use case and propose the optimal solution. Order the development of an AI market making bot and get a consultation on your strategy.