At a Rise of AI Virtual Chat recorded in July 2020, Fabian Westerheide interviews trader Raul Glavan, who describes building an AI system that automates and, he says, now outperforms his own decades of manual trading. Glavan traces his path from childhood chess and an early home computer through winning Germany's Iron Trader competition and the Forex World Cup to a system trained specifically on his own trading behavior rather than generic market data. He explains a mean-reversion strategy grounded in behavioral finance, buying into panic and selling into euphoria across stocks, commodities and currencies, using scraped sentiment data and LSTM neural networks with peephole connections built in Python and TensorFlow. He reports around five trades a week held roughly 30 days each, a 93 percent historical win rate, a profit factor of 24, and a Sharpe ratio near 20, which he calls unsustainable given his small sample size. He also describes overruling his own system during the March 2020 Corona panic and losing money as a result. The conversation closes with a discussion of universal basic income and equal taxation of AI and human income as responses to automation-driven job loss.
Key takeaways
- Glavan built his AI trading system specifically to replicate and improve on his own trading behavior rather than training on generic market data.
- His mean-reversion strategy buys into market panic and sells into euphoria across stocks, commodities and currencies.
- He uses LSTM networks with peephole connections, built in Python and TensorFlow, and combines them with scraped social media and insider-trading signals.
- He reports a roughly 93 percent win rate and a Sharpe ratio near 20, but calls both figures unsustainable given his small sample size.
- He and Westerheide argue AI-driven and capital income should be taxed the same as labor income to help fund universal basic income.
Detailed notes
From chess and a home computer to algorithmic trading
Glavan describes starting chess at age six, getting an early home computer as a present and teaching himself to program on it, and making his first real stock trades as a teenager on gut feeling before building his first rule-based trend-following trading system at 16. As a student he built an early robo-advisor, though its exact name is unclear from the transcript, and won Germany's Iron Trader competition and later the Forex World Cup, where he says he generated returns of several hundred percent virtually over a few months using high leverage.
A mean-reversion strategy built on behavioral finance
Glavan says his strategy exploits behavioral biases such as panic and euphoria: when other traders are panic-selling, his system buys, and vice versa, a move he calls counter-intuitive because it goes against instinct. He scans stocks, commodities and currencies rather than single stocks, citing the roughly eight trillion dollars traded daily in foreign exchange, and avoids concentrated single-stock bets because of bankruptcy risk, citing Wirecard's collapse to zero as an example of a stock that kept falling after an initial panic signal.
Sentiment data and the model behind the system
Glavan says he began scraping Twitter for trading signals in 2013, moving from simple keyword spotting toward natural language processing that can handle negation, and also tracks high-influence accounts and publicly disclosed insider trades, weighting a chief executive's stock purchases against their known salary to judge how meaningful the trade is. His model uses LSTM networks with peephole connections, built in Python with TensorFlow, which he found more accurate for forecasting than the alternatives he tested. He runs the system on his own physical computer rather than the cloud for security, and trades about five times a week with roughly 30-day holding periods.
Performance, risk and a costly human override
Glavan reports a roughly 93 percent historical win rate, a profit factor of 24, and a Sharpe ratio near 20 for the period discussed, figures he calls insane and not sustainable given his limited sample size and the difficulty of beating an efficient market like currencies. He does not use stop-losses, arguing they get exploited by market makers in the illiquid options he trades, relying instead on diversified position sizing. He also describes panicking during the March 2020 Corona sell-off and manually reducing his position sizes instead of following his system's signals, which he says cost him money he would otherwise have made.
UBI and equal taxation of AI and human income
Westerheide raises concern that AI concentrates wealth and power, and Glavan agrees that automation and digitalization risk mass unemployment, arguing for universal basic income guaranteeing free healthcare, education and housing so people can take risks. Citing an Oxford paper on large tech companies' low effective tax rates, Glavan and Westerheide argue capital and AI-driven earnings should be taxed the same as labor income; a live audience poll cited during the discussion found 75 percent in favor of taxing AI and human earnings equally, 8 percent wanting AI taxed higher, and 17 percent wanting it taxed lower.
Quotes
After now something like 20 years I recognized I think it's the hardest way to make easy money.
Raul Glavan
My system is now better than me, so far better.
Raul Glavan
Every entrepreneur out there needs his own UBI.
Fabian Westerheide
Questions answered in this talk
How does Raul Glavan's AI trading system decide when to buy?
According to Raul Glavan at a Rise of AI Virtual Chat in 2020, his system looks for panic or euphoria across stocks, commodities and currencies using scraped sentiment and historical trading patterns, and buys into panic and sells into euphoria based on behavioral finance.
What win rate and Sharpe ratio does Raul Glavan report for his trading system?
Raul Glavan said at a Rise of AI Virtual Chat in 2020 that his system had a roughly 93 percent historical win rate, a profit factor of 24, and a Sharpe ratio near 20 for that period, which he called unsustainable given his small sample size.
What happened when Raul Glavan overrode his own AI trading system during the Corona market panic?
According to Raul Glavan at a Rise of AI Virtual Chat in 2020, he panicked in March 2020 and manually reduced his position sizes instead of following his system's signals, which he says cost him money he would otherwise have made.