Trader Miles Deutscher implemented an automated trading strategy based on the GARCH model using the Claude Code tool. According to him, the entire system was assembled by artificial intelligence — the human only set the task.

The GARCH (Generalized Autoregressive Conditional Heteroskedasticity) model is one of the most powerful quantitative tools used daily by professional quantitative analysts. Its predecessor, the ARCH model, was developed by Robert Engle in 2003 and awarded the Nobel Prize. Later, Engle's student, Tim Bollerslev, expanded it into GARCH, which is still used in quantitative analysis departments of leading investment houses.

How the model works and what it measures

The main idea of GARCH is that market volatility is not chaotic but can be predicted. Deutscher explained that the model allows measuring and forecasting fluctuations in advance, relying on just three indicators.

First — the baseline level of volatility that every asset has and that hardly changes. For Bitcoin, this level is traditionally higher than for the S&P 500 index. Second — the strength of yesterday's market movement. Third — the volatility level at which the market was before that movement. The sum of these three values provides a forecast of fluctuations for the next day.

AI played a key role in the project. Deutscher emphasized that he did not program the system himself but described the task and entrusted the assembly to Claude Code, which turned the model's mathematics into a ready-to-use working tool.

An important limitation: the model shows how much the market can move but does not indicate the direction. This is a volatility forecast for portfolio risk management, not a buy or sell signal.

From forecast to position size

Deutscher uses the obtained fluctuation estimate to calculate the position size. Institutional departments typically limit annual drawdown to 15%, and all subsequent mathematics is built from this limit.

Example: if the fluctuation estimate for an asset is 30% annually, the drawdown limit is divided by it — and the position multiplier becomes 0.5x. With a base position of $10,000, this means an entry of $5,000.

This approach does not predict price movement but only controls exposure to already measured risk. For clarity, Deutscher added that at the time of writing, the fluctuation estimate for the Nasdaq index stood at 30% annually.

Backtesting on historical data

The trader tested the model's real utility on a fifteen-year dataset for Bitcoin. He ran the same strategy on a moving average crossover in two ways — with identical entries, exits, and signals.

With a fixed position size, the final capital in the test was $17,957. With position sizing based on the GARCH model, the result was higher — $21,205. According to Deutscher's assessment, the version with the Nobel-winning model not only generated more profit but also navigated the path with smaller drawdowns and overall risk.

However, the expert warned of a significant limitation: during a sharp crash following a calm period, the model will not work, and the position will be opened in full size.

Analyst's opinion: Integrating classical econometric models with modern AI agents is a natural stage in the evolution of trading. GARCH remains the gold standard for risk assessment, and automation through Claude Code makes it accessible not only to institutional but also to retail traders. However, blindly trusting the model without understanding its limitations is a direct path to losses during market shocks.