At the Rise of AI Summit 2020, Raul Glavan, a quant trader speaking at his fifth Rise of AI event, explains how he builds and runs AI trading models with his own capital. He quit his job in asset management to trade full time and entered the World Championship of Forex Trading, reporting about 150% net return and a 2.3% maximum drawdown for the year at the time of the talk. He argues that building a profitable AI trading system is possible but very hard: easily accessible public data such as prices and technical indicators is highly competitive and has little predictive value, so an edge requires accurate alternative data, rigorous feature engineering, and avoiding errors like look-ahead bias. He walks through his six-step process, data sources ranging from sentiment analysis to insider trading ratios and unconventional signals like sudden LinkedIn profile changes that hint at upcoming layoffs, and his practice of building a separate model per asset. He closes with audience questions on market efficiency, data infrastructure, time allocation and advice for beginners.
Key takeaways
- Glavan quit his asset management job to trade his own capital with AI, reporting about 150% net return in forex trading that year.
- He argues easily accessible public price and technical-indicator data has little predictive value; alternative data like sentiment matters more.
- He builds a separate model per asset, since one model cannot generalize across assets like stocks, copper and currency pairs.
- Look-ahead bias, such as normalizing training and test data together, is a common and dangerous error he sees in many trading models.
- He spends about 80% of his time on data acquisition and feature engineering, only 10% on the model itself.
Detailed notes
From asset management to trading his own capital
Glavan says he quit his job in the asset management industry the previous year to trade his own capital using AI and entered the public World Championship of Forex Trading to prove the approach works. At the time of the talk he reports being in first place for most weeks that year, with about 150% net return and a 2.3% maximum drawdown, verified daily by a third party.
Why a money printing machine is hard to build
He says building profitable trading models is possible, as quant funds, prop shops and hedge funds show, but genuinely difficult. Public data such as stock and asset prices is cheap, widely used and has almost no prognostic value on its own. Useful alternative or proprietary data is hard and expensive to obtain and often unstructured. Markets are highly competitive and efficient, transaction costs erode edges, and countless variables influence prices, making sustained advantage rare.
A practical process and common mistakes
Glavan outlines a six-step process, saying data acquisition and feature engineering take about 80% of his time. He warns against using technical indicators as core features, since they are just derivatives of price and often collinear, which slows models and reduces predictive value. He favors LSTMs for time series and decision trees for their easier feature-importance analysis. He repeatedly flags look-ahead bias as a common, dangerous error, for example normalizing training and test data together, or using economic data published later than the point a model is meant to predict from.
From easy data to hard and unstructured data
He distinguishes 'easy' public data, fundamentals, prices and technical indicators, which most traders already use and which has little value, from 'hard' public data such as sentiment analysis and insider trading. For insider trading he stresses using a person's track record and the ratio of a trade's size to their income, not the raw fact that an insider bought stock. He gives crawling LinkedIn for sudden employee profile changes as a signal of an impending layoff as one example of building an information edge from public but unstructured sources, and lists patent data, credit card transactions, IoT and weather data as further unstructured private data sources.
Model design and results
Glavan builds a separate model for each asset, saying a single model cannot meaningfully forecast an Amazon stock, the copper price and a currency pair alike. He shows feature-importance examples for an S&P model, dominated by corporate insider data and liquidity metrics, and for a copper model, dominated by supply and demand factors, along with recent buy and sell signals on a roughly four-week horizon. He contrasts an implausible back-tested claim of 90% accuracy and a profit factor over 30 with his own reported real-money results of 94% accuracy this year.
Q&A highlights
Asked what happens as more traders use AI models, Glavan says competition will rise, markets will become more efficient, spreads will narrow and profits will shrink, which he calls good for the market overall but harder for model builders. On infrastructure, he says he avoids cloud services and stores all data on his own servers out of concern for protecting his code and data. He estimates his time as 80% feature engineering, 10% modeling and 10% idea generation, says he relies on a network of domain experts such as weather specialists for commodities like orange juice, and advises beginners to learn Python and libraries like TensorFlow and Keras, then replicate and modify simple public models on GitHub.
Quotes
I would say feature engineering is 80%, model is 10%, and the other 10% is something like idea generation.
Raul Glavan
Markets are very efficient, and it's very hard to gain an edge against other market participants.
Raul Glavan
Questions answered in this talk
Is it possible to build a profitable AI trading system?
According to Raul Glavan at Rise of AI Summit 2020, it is possible but very hard, because public price data is highly competitive with little predictive value, so success depends on accurate alternative data, careful feature engineering and avoiding errors like look-ahead bias.
What data sources give an edge in AI-driven trading?
According to Raul Glavan at Rise of AI Summit 2020, sentiment analysis, insider trading data combined with a trader's track record, and unconventional sources such as LinkedIn hiring changes, patents or credit card data can provide more predictive value than easily accessible price and technical-indicator data.
How much time should go into building an AI trading model?
According to Raul Glavan at Rise of AI Summit 2020, about 80% of the effort should go into data acquisition and feature engineering, 10% into building the model itself, and 10% into generating new ideas.