Speaker
Grigory Chikishev, Team Lead and Quantitative Trader at Quantum Brains with over 9 years of experience.
The growing interest in AI-driven trading is fueled by headlines that often appear to signal a breakthrough. In February 2025, Bloomberg reported that Minotaur Capital — a fund partially relying on algorithmic decision-making — delivered a 14% return in its first six months, outperforming the global equity index, which rose roughly 7% during the same period. Such numbers attract attention and create the impression of a systematic edge.
Yet short-term outperformance alone does not prove long-term viability. Temporary gains may reflect luck or market conditions rather than model quality. As numerous experiments have shown, even simple or random strategies can occasionally beat human traders over limited periods.
AI as a Mathematical System, Not Market Intuition
Modern AI systems are built on statistical methods and large datasets — not intuition or human-like insight. These models excel where there are clear, repetitive patterns and abundant labeled data: image classification, speech recognition, text analysis, anomaly detection.
Financial markets, however, are dominated by noise. Price movements reflect countless interacting factors, and the true predictive signal is extremely weak. Effective use of AI therefore requires rigorous data preparation, including:
- Averaging and smoothing inputs (e.g., using volume-weighted mid-prices instead of raw last-trade prices to reduce noise).
- Aggregating or redefining prediction targets (e.g., forecasting average price over a period rather than a single tick or second).
- Reducing microstructure noise through windowing, collapsing ultra-short-term fluctuations, and focusing on stable metrics.
These techniques expose a cleaner structure that machine-learning models can realistically learn, aligning the prediction problem with what statistical algorithms are designed to handle.
Data Scarcity and Ways to Expand Training Sets
Market history is limited, especially at higher frequencies. A full year of price data may yield fewer than 10 million observations — insufficient for modern high-capacity models. To address this, practitioners rely on:
- Bootstrapping, which generates multiple sub-samples for more robust statistical estimates;
- Time-series augmentation, such as scaling or stretching, that expands the dataset without breaking temporal relationships.
These methods help create richer training environments without distorting market dynamics.
Extreme Risks and the Challenge of “Black Swan” Events
Markets occasionally experience rare, sudden, high-impact disruptions: liquidity collapses, unexpected volatility spikes, systemic shocks. These are commonly referred to as “black swans” — events that occur infrequently yet reshape entire markets.
Because such events may happen once in a decade, there is simply not enough data for a model to meaningfully “learn” them. AI cannot generalize from what it has never seen. Managing these risks falls squarely within risk management, not prediction. AI can support early anomaly detection, but cannot guarantee resilience against such shocks — much like monitoring for meteorites cannot prevent an impact.
Where AI Clearly Adds Value
AI delivers reliable advantages in areas where results can be verified before execution or where the data depth is significant:
- sentiment analysis across news and social platforms;
- identifying arbitrage opportunities;
- market-making and order-book analytics;
- automation of complex, data-heavy operational processes.
Recommendations
- Apply AI to problems that can be clearly formalized and validated.
- Avoid relying on models that attempt to predict price direction in low-signal environments.
- Combine any AI-driven system with rigorous risk management.
- For fintech startups, focus on data infrastructure, analytics, anomaly detection, and execution tooling rather than building “models that beat the market.” These areas demand less historical data, involve lower risk, and result in faster, more dependable products.
Conclusion
AI is a powerful enabler in trading and market analytics, but its effectiveness is defined by the nature of financial data. It performs best where patterns are stable and outcomes can be verified — not in predicting inherently noisy price movements. When applied with discipline and supported by sound risk controls, AI enhances decision-making precisely because it is mathematical and emotion-free — a valuable counterbalance in an industry where human emotions often cause greater losses than model errors.