When applying machine learning to noisy financial data, the whole system's design is as important as the development itself. I design the fundamental idea and transform it into a working algorithm. Adding an intelligent layer in the form of artificial intelligence will give it an advantage and deliver an uncorrelated strategy that will be a vital addition to your portfolio.
Peter Kostovcik is a professional quantitative researcher and co-founder of an AI algorithmic fund in Prague, Czech Republic. He lives in Medellín, Colombia, and connects a lifestyle with a job he loves.
This strategy exploits statistical biases mainly in equity markets. The portfolio is re-balanced daily to respond ever-changing market environment. The system is long only, but it trades the whole universe of assets. Thanks to ETFs, it may enter short positions (inverse ETFs) or invest in non-equity markets such as commodities, bonds, etc. The idea behind the bias has a fundamental explanation and was empirically observed. The exceptional edge in this strategy holds a neural network implementation in the final decision process (daily up to 30 positions in different stocks/ETFs and around 5 different future contracts).
Actual asset universe: US stocks, ETFs, and futures.
Expected returns based on historical performance: 15-20% p.a.
Expected max drawdowns are around 20%, up to 35% by Monte Carlo analysis.
Well-known strategy in the industry that enters each trade into long and short positions, so it exploits the mean-reversion relationship. Still, the system stays neutral to the market as a whole. This system uses auto-encoders based on convolutional neural networks to find stock pairs with mean-reversion relationships. Stocks are from the same industry or have a customer-supplier relationship. Next to classical approaches such as co-integration and distance method, the auto-encoders proved to be a valid alternative.
The selected universe of stock pairs forms thousands of potential signals, so here comes graph neural networks to select the best signals according to expected returns. Advanced feature engineering and a proprietary cross-validation approach to time series are vital to the development process.
Expected returns: 12-15% p.a.
Expected MDD around 17% (30% by Monte Carlo)