Backtesting market making strategies in crypto landscapes

This research evaluates two market making models—one static (Roll-based) and one adaptive (GARCH-based, using dynamic predictions of conditional volatility)—within a custom-built simulation environment using high-frequency, real-world time-series data (cryptocurrency trades and order book snapshots). The study compares their performance by backtesting their trades signaled by them on an identical historical data stream of trades of a CLANKER, which is a cryptocurrency from Bybit. The adaptive GARCH model, which dynamically adjusted its conditional volatility predictions based on a GARCH(1,1) model derived from AR(1) residuals, exhibited significantly better risk management. Whereas, the Roll model’s assumed homoskedasticity and used a constant spread for the bid and ask quotes. The results displayed a nearly identical Sharpe ratios for both strategies; however, with the GARCH model yielding a significantly lower portfolio variance and inventory risk. This was because both strategies ended up accumulating net long positions throughout the backtest. However, GARCH model’s inventory accumulation was much lower than Roll model (because of it’s tight spread causing more trades). While the simpler Roll model had a higher interaction rate (i.e., trade count), it’s lack of adaptivity to real time volatility led to higher risk exposure. These findings empirically suggest that incorporating dynamic, data-driven features enhances model robustness and risk control in processing high-velocity data streams, especially for high frequency trading and market making.