Developing an ML-Based Arbitrage Trading Bot

Developing an ML-Based Arbitrage Trading Bot

Pure arbitrage (risk-free profit) was killed long ago by HFT algorithms. Real money lies in statistical strategies that carry risk, and that's where ML becomes indispensable. We build such bots from scratch: from collecting tick-level order books to deploying on co-located servers near exchanges. In one project, we built a pairs trading system that has maintained a Sharpe ratio above 2.0 for over two years. Let's start with the main point: how does ML find arbitrage opportunities? It requires cointegration—a statistical relationship between the prices of two assets. When the spread deviates from equilibrium, we open a position. But for the spread to be stable, proper pair selection and an adaptive hedge ratio are needed. Our stack: Python for prototyping, C++ for execution. Below we break down the details—tables, code, and a step-by-step guide.

How ML Finds Arbitrage Pairs

We start with 500+ assets, compute a correlation matrix, cluster them (K-means), and select pairs within clusters. We test for cointegration using the Engle–Granger test. From ~125,000 pairs, 50–200 tradeable ones remain. Then for each pair we calculate a dynamic hedge ratio via Kalman Filter. Below is production code from a project on Deribit.

import numpy as np
from statsmodels.tsa.stattools import coint
from pykalman import KalmanFilter

# Kalman Filter for dynamic hedge ratio
kf = KalmanFilter(
    transition_matrices=[1],
    observation_matrices=[1],
    initial_state_mean=0,
    initial_state_covariance=1,
    observation_covariance=0.5,
    transition_covariance=0.1
)

# hedge_ratios = state means over time
state_means, _ = kf.filter(spreads.values)

Why Kalman Filter? It adapts to market changes, unlike static OLS. Comparison table:

| Parameter | OLS | Kalman Filter | | Adaptivity | Static | Dynamic | | Sharpe ratio | 1.2 | 1.8 | | Maximum drawdown | 15% | 8% | | Latency sensitivity | No | No |

The Kalman Filter yields a Sharpe 50% higher than OLS.

Types of Arbitrage Strategies: Table

| Type | Example | Risks | ML Task | Expected Return | | Cross-exchange | BTC/USD Binance vs Bybit | Latency, fees | Spread persistence prediction | 0.1–0.5% per trade | | Pairs trading | ETH/BTC | Cointegration breakdown | Pair selection, hedge ratio | 15–25% annual | | Triangular | BTC/USDT → ETH/BTC → ETH/USDT | Fees | Path ranking | 0.05–0.2% per cycle | | ETF | SPY vs NAV | Liquidity | Divergence prediction | 5–10% annual | | Funding rate | Spot + perp | Trend reversal | Entry/exit signal | 10–30% annual |

Why Latency Is Critical for Cross-Exchange Arbitrage

At high frequencies, the window is a few milliseconds. We place resting orders on both exchanges and upon detecting a discrepancy simultaneously cancel and place new ones. Co-location in data centers (Equinix NY4/LD4) with kernel bypass (DPDK) and FPGA for ultra-low latency. In one project we achieved p99 < 500 µs.

For stat arb with a daily horizon, latency is less important—a VPS suffices.

How to Build a Pairs Trading Bot: Step by Step

1. Data collection — tick-level order books via WebSocket, stored in ClickHouse.
2. Clustering — correlation matrix → K-means.
3. Pair selection — cointegration test, p-value < 0.05.
4. Hedge ratio calculation — Kalman Filter (dynamic) or OLS (static).
5. Risk management — Value-at-Risk, stop-loss on spread.
6. Backtesting — on historical data with slippage and fees.
7. Deployment — Docker containers, Gitlab CI, monitoring with Prometheus + Grafana.

Signal Decay: How Not to Lose Money

Statistical strategies degrade—alpha dissipates. We monitor each pair's rolling Sharpe ratio. If it drops below 1.0 over 30 trading days, the pair is removed and the model retrained.

What's Included in the Work

• Strategy and architecture documentation
• REST API for broker/exchange integration
• Source code in Python/C++ with unit tests
• Deployment instructions (Docker, Ansible)
• Client team training (2–3 sessions)
• 3 months of technical support

We guarantee code quality and trade security.

Our team has over 10 years of experience in HFT and ML. We have completed 25+ algorithmic trading projects. We work with crypto exchanges, brokers, and funds.

Contact us to evaluate your project and propose an optimal architecture. Get a consultation on strategy and stack selection.

Industry AI Solutions: Healthcare, Finance, Retail, Manufacturing

We encounter the same pain points: a general text model doesn’t distinguish medical nomenclature, and a standard object detector confuses “weld seam scratch” with “casing scratch.” Each time these are different defects with different consequences. To avoid this, we build industry-specific solutions on top of general methods, but with deep domain knowledge — from regulatory requirements to data specifics. Over 5 years, we have completed 80+ projects in fintech, healthcare, retail, and manufacturing, and none were without adaptation to a specific business case.

Healthcare: Regulatory Maze and Data Governance

Medical AI differs not in technical algorithms but in a compliance-first approach. Depending on the country of application, the model may be a Class II or III medical device requiring clinical trials (FDA, CE MDR, GOST R). We ensure compliance with these standards at the architecture stage — fixing them post-factum is 10× more expensive.

Medical imaging. Detection on X‑rays, CT, MRI is a mature area. Models on ResNet, EfficientNet, SegFormer achieve AUC 0.94–0.97 on standard tasks (pneumonia on CXR, polyps on colonoscopy). Key issue is generalization: a model trained on data from one scanner manufacturer degrades on another due to differences in preprocessing and artifacts. Solution: domain adaptation via MONAI (Medical Open Network for AI) from NVIDIA, which includes DICOM loading, 3D augmentation, and confidence calibration. TotalSegmentator — for automatic segmentation of 117 structures on CT, production‑ready, Apache 2.0 license.

Clinical NLP. Extracting structured information from clinical records: diagnoses (ICD‑10/11), prescriptions, dates, indicators. medspaCy, scispaCy, MedCAT — specialized NLP libraries with ontologies (SNOMED‑CT, UMLS). Fine‑tuning BioBERT or ClinicalBERT on our data yields F1 0.85–0.92 on NER tasks versus F1 0.65–0.72 for general BERT. We verified this on a project with a regional oncology center — cancer stage extraction accuracy increased by 23%.

Clinical decision support. LLM assistants for clinical decision support are a regulatory gray area. We use an RAG system on top of clinical guidelines (UpToDate, local protocols) with explicit citation for each statement. The model does not diagnose but helps find relevant protocols. Stack: LlamaIndex + pgvector + pubmedbert-base-embeddings + Llama Guard for safety. Data in DICOM/HL7 FHIR, on‑premise deployment mandatory.

Finance: How to Ensure Interpretability of a Scoring Model under Basel IV?

The financial sector is one of the most mature in applying ML, but regulation is maximal. Every model affecting credit decisions falls under Basel IV, EU AI Act, GDPR Article 22. We deliver AI solutions for fintech that satisfy these requirements — in a project for a top‑10 bank we deployed a scoring model where each record required SHAP explanations.

Credit scoring. Gradient boosting (LightGBM, XGBoost) dominates. Neural networks yield +0.5–2% AUC but lose interpretability. Standard: LightGBM + SHAP to explain each decision. Fairness checking is mandatory: Fairlearn or aif360 for auditing disparate impact on protected attributes (age, gender). The default class is 1–5% — with an imbalance of 1:30, a model with 97% accuracy may have recall 0.2. Solution: focal loss, class_weight='balanced', SMOTE + careful validation. In one fintech scoring project, the model reduced credit losses by $2.1 million annually.

Algorithmic trading and risk management. LSTM and Transformer for price forecasting are popular but unstable in production due to non‑stationarity of financial series. A more robust approach: ML for signal generation (classification: up/down over horizon N) with traditional portfolio optimization on top. Backtesting via Zipline‑Reloaded, vectorbt, QuantLib. Proper backtesting is critical — look‑ahead bias kills results. We guarantee a clean experiment: all data at signal time is available in real time.

AML (Anti‑Money Laundering). Graph Neural Networks for analyzing transaction networks is an actively developing area. PyG, DGL for GNN. Task: detect suspicious patterns in transaction graphs (layering, structuring). Recall is more critical than precision — better 10 false alarms than miss one money laundering. In a project for a large payment service, we increased recall by 18% without increasing false positive rate.

Retail and e‑commerce: Recommendation Systems and Demand Forecasting

Recommendation systems. Current architectural standard: two‑tower model for retrieval + ranking with cross‑features. TensorFlow Recommenders or Merlin from NVIDIA for GPU‑accelerated feature processing. For small catalogs (<100k items), LightFM is sufficient. A common mistake is training on implicit feedback without accounting for position bias. Solution: IPW (Inverse Propensity Weighting) or randomized logging on a portion of traffic. Development time for a basic recommendation system is 4–8 weeks, including A/B test.

Demand forecasting and inventory optimization. Hierarchical forecasting: SKU → category → store → region. HierarchicalForecast from Nixtla automatically reconciles forecasts across levels. TFT or N‑HiTS for base forecast, gradient boosting for adjustment on exogenous factors (promotions, weather, events). One retail project led to a 15% reduction in stock‑outs due to precise promotion calibration.

Visual search and size compatibility. CLIP embeddings for image search — deploy in 2–3 weeks: clip‑ViT‑B‑32 or clip‑ViT‑L‑14, Faiss or Qdrant index, REST API. For size recommendation — specific models on return data and reviews with fit indication.

Manufacturing: Quality Inspection and Predictive Maintenance

Quality control and defect detection. CV models for product inspection are one of the most mature industry tasks. YOLOv10 for defect detection, SegFormer for segmentation. Specifics: class imbalance (defects are rare), high recall requirement (missing a defect is worse than false alarm). Typical dataset: 500–2000 defect images + 500–1000 normal. Few‑shot learning via DINO or SAM 2 works with 50–100 annotated examples. We gained experience on an electronics production line — recall 0.95 at FPR 0.03. A predictive maintenance deployment saved a manufacturing client $500,000 per year in unplanned downtime.

Predictive maintenance. Vibration sensors, current sensors, thermocouples → feature extraction → anomaly or mode classification. Models: LSTM‑AE for unsupervised, LightGBM for supervised (if failure history is available). Integration with SCADA/OPC‑UA via opcua-asyncio or MQTT. Key metric: False Negative Rate — a missed pre‑failure is more costly than a false alarm. Threshold tuned to business cost of each error type. Timeline: 3 to 6 months to production.

Digital twin and simulation. Surrogate models — ML models replacing expensive physical simulation. If a CFD simulation takes 6 hours and a surrogate (trained on 10,000 simulations) takes 0.01 seconds, that's 2,000,000× speedup for optimization. SALib for sensitivity analysis, botorch for Bayesian optimization on top of surrogate.

General Principles of Industry AI

Regardless of industry, there are patterns that work everywhere. Data matters more than architecture. In healthcare, 1000 quality labeled images are better than 100,000 poor ones. In manufacturing, 200 real defect examples are more valuable than 10,000 synthetic ones. Compliance‑first design — regulatory requirements are easier to embed into architecture from the start than to add later. Logging, explainability, versioning from day one. Domain expert on the team — an ML engineer without domain knowledge does slowly and error‑prone what an ML engineer plus a doctor/financier/technologist does quickly and correctly.

We guarantee certification to customer requirements (ISO 13485, SOC 2, GDPR) and provide full model documentation (model card, datasheet, compliance report). Our experience: 10,000+ engineering hours and 80+ projects.

Work Process for an Industry AI Solution

1. Domain immersion (2–3 days) — interviews with experts, studying regulatory requirements, auditing available data.
2. MVP design (1–2 weeks) — stack and architecture selection, feasibility assessment.
3. Development and validation (from 4 weeks to 6 months depending on industry) — model training, testing, compliance.
4. Integration and deployment (1–4 weeks) — on‑premise / cloud / edge, documentation, staff training.
5. Support and monitoring — model drift, retraining, SLA.

Estimated timelines:

Cost is calculated individually for each project. Get a consultation — we will evaluate your dataset, regulatory map, and business goals.

Why Choose Our Industry AI Solutions?

• 80+ completed projects in fintech, healthcare, retail, and manufacturing.
• 5 years on the market — proven experience with compliance and deployment.
• Quality guarantee: we ensure target metrics (AUC, recall, latency p99) and provide full documentation.
• Licensed technologies: PyTorch, MONAI, LightGBM, Qdrant — we use open‑source with commercially safe licenses.
• Flexibility: we work as a contractor or as an extension of your team.

Contact us for a free data audit and consultation. Request a proposal with a detailed work plan. We will discuss your task and prepare a commercial proposal.