AI Based Real Time detection of Blockchain Market Anomalies

Abstract

This research proposes an artificial intelligence (AI)-based real-time anomaly detection system that perceives anomalies in Bitcoin perpetual futures markets. The limitation of traditional static monitoring tools resolved in the proposed framework as static thresholds cannot differentiate between a true anomaly or a candle that does not strictly fall within a given range. In trending stable market conditions, a threshold-based technique fires on normal candles resulting in alert fatigue instead of actionable signals.The proposed monitoring system makes use of the Isolation Forest algorithm, which is an unsupervised machine learning algorithm that detects anomalous behaviour in price and volume time series without any labelled training data. The architecture consists of three coupled but functionally independent analytical modules: Price/Volume Module featuring Isolation Forest and Order Flow Imbalance analysis, Liquidation Heatmap Proxy based on open interest public data, and Funding Rate Module based on perpetual future funding mechanism. Each module generates an independent directional probability estimate on a scale. The final UP/DOWN/NEUTRAL signal model is a weighted average (with weights easily verifiable) of sub-module signals. The model is based solely on data obtained from the Binance Futures public REST API documenting the reproducibility of the model. The prototype was demonstrated with a live simulation for a period of two weeks using the BTCUSDT perpetual futures pair. It was noted that the Isolation Forest model demonstrated the ability to adapt to changing volatility regimes, the Order Flow Imbalance indicator was capable of providing responsive trend indications, and significant events such as a spike in volume, extreme funding rates and rise in volatility were also successfully detected. During live testing over the course of two weeks, the system has successfully discerned three distinct types of events: identified volume spike with no price movement, identified extreme funding rate in conjunction with price heatmap bearish imbalance, and identified “Anomalous” volatility regime with OFI at +0.11 — these events would not have been able to be successfully identified in a fixed-threshold machine learning system without the confounding occurrence of many false-positive events.The full source code for the system proposed in this thesis is publicly available in GitHub. The repository includes all detection modules, the developed Streamlit dashboard and requirements.txt fulfilling the requirements to create the same running environment. The system has no secret API keys, paid services or additional computing power required.