Trader Miles Deutscher presented a unique development: he fully automated a trading strategy based on the GARCH model using the Claude Code tool. Notably, the entire system was built by artificial intelligence — the trader only described the task, and the neural network turned complex mathematics into a working tool.

What is GARCH and why it matters

The GARCH model (Generalized Autoregressive Conditional Heteroskedasticity) is one of the most powerful quantitative models that professional quants use daily. Its predecessor, ARCH, was developed by Robert Engle, who was awarded the Nobel Prize in 2003, and then his student Tim Bollerslev expanded it into GARCH. The key idea: market volatility is not random — it is predictable.

The model relies on just three indicators:

  • Base level of fluctuations — a constant for each asset. For Bitcoin, it is traditionally higher than for the S&P 500 index.
  • Strength of yesterday's movement — how much the market changed over the last day.
  • Previous volatility level — the level of fluctuations the market was at before this movement.

The sum of these three values gives a volatility forecast for tomorrow. Deutscher emphasizes: this forecast does not indicate the direction of movement, but only measures how much the market might swing. It is a risk management tool, not a buy or sell signal.

How the forecast turns into position size

The trader uses the obtained fluctuation estimate to calculate position size. The standard approach of institutional departments: limit the annual drawdown to 15%. Then the limit is divided by the forecasted volatility. For example, if the fluctuation estimate for an asset is 30% annually, the position multiplier becomes 0.5x. With a base position of $10,000, this means entering at $5,000.

Deutscher tested the strategy on a fifteen-year dataset for Bitcoin, comparing two approaches: with a fixed position size and with calculation based on the GARCH model. The results are impressive: with a fixed size, the final capital was $17,957, while with GARCH it was $21,205. Moreover, the second option went through with smaller drawdowns and overall risk.

Limitations and risks

The trader honestly warned about the main limitation: the model does not work during a sharp crash after a calm period. In such a scenario, the position will be opened in full, which could lead to significant losses. GARCH is powerful but not a panacea. It is good for risk management in normal market conditions but does not protect against "black swans."

Expert opinion: Automating GARCH through AI is a logical step in the evolution of trading. The Nobel Prize-winning model becomes accessible not only to quants but also to retail traders. However, it is important to remember: any algorithm, even the most advanced, requires an understanding of its limitations and manual control in extreme situations.