Can AI Actually Help You Trade? What the Research Shows

What the Academic Research Says

The academic literature on AI-assisted trading is growing but comes with significant caveats. Most studies are backtests, not live trading results. The gap between the two is where most retail traders get hurt.

The Zarattini et al. 2024 Study: 36% Alpha (With Caveats)

One of the most widely cited recent papers is Zarattini, Aziz, Buriro, and Karimov (2024), which applied a systematic Opening Range Breakout (ORB) strategy with machine learning filters. Their backtest showed 36% annualized alpha over 27 years of data.

That’s an extraordinary claim, and it deserves scrutiny:

• This is a backtest, not live trading. The strategy was tested on historical data, not executed in real markets with real slippage, liquidity constraints, and execution costs. Backtested strategies routinely degrade 30–50% or more when deployed live.
• Conflict of interest disclosure: Co-author Andrew Aziz is the founder of Bear Bull Traders, a day trading education company. He has a commercial interest in demonstrating that systematic day trading can work. This does not invalidate the research, but it is relevant context.
• The strategy is highly systematic. It is not “AI picks stocks for you.” It is a rules-based system with strict entry/exit criteria, position sizing, and risk management. The ML component filters which setups to take — it does not replace discipline.

The Zarattini study is interesting because it suggests that systematic, rules-based approaches can generate alpha — a finding consistent with the broader quantitative finance literature. But it does not demonstrate that AI makes discretionary day traders profitable. Those are very different claims.

The Broader ML Literature: Promising but Fragile

Dozens of papers have applied machine learning to stock prediction. LSTM neural networks, gradient-boosted trees, sentiment analysis of news and social media, reinforcement learning for portfolio allocation — the approaches vary widely. Common findings across the literature:

• In-sample performance is almost always impressive. ML models can fit historical patterns extremely well. This is often mistaken for predictive power.
• Out-of-sample performance degrades significantly. Overfitting is the central challenge. Models that work on training data frequently fail on new data.
• Transaction costs matter enormously. Many papers report gross returns. Once you add realistic spreads, commissions, and slippage, strategies that looked profitable become losers. The Beckmeyer et al. (2023) study of 0DTE options found that 60% of retail losses came from transaction costs alone.
• Shorter horizons = more noise. The signal-to-noise ratio in financial data drops dramatically at intraday and daily frequencies. ML models have more to work with at weekly, monthly, and quarterly horizons.

What AI Can Actually Do: Systematic Signals at Longer Horizons

AI will not make you a successful day trader. But systematic signal-following at longer horizons tells a different story.

The consistent finding across the academic literature is that quantitative, rules-based strategies work best when they:

1. Operate on medium-to-long time horizons (weeks to months, not minutes to hours)
2. Use multiple signal types (price momentum, volume, sentiment, fundamentals) rather than any single indicator
3. Apply strict, predefined rules for entry, exit, and position sizing
4. Account for realistic transaction costs in their performance measurement

This is where our own data becomes relevant.

Original Data: 15.1 Million Scored Stock-Days

Banana Farmer’s scoring system combines technical momentum with social sentiment data to rank ~9,000 stocks and crypto assets in real time. We backtested this system across 15.1 million scored stock-days (12 years of data, ~5,000 assets) and validated against 5,900+ live production signals.

The results directly illustrate the horizon effect:

The pattern is unambiguous. At day-trading horizons (1–5 days), even systematically generated signals produce coin-flip results — 47–50% win rates. This is consistent with everything the academic literature says about short-term trading.

But at 60 days, backtested “Ripe”-rated signals hit a 57.7% win rate with a median return of +2.09%. At 3 months, live production signals confirmed the pattern: 58.5% win rate across 5,547 measured signals. The backtest and live data converge, suggesting the signal is real and not overfit.

Critically, “Ripening” signals (stocks building momentum) outperformed “Ripe” signals (peak momentum) at 60 days: 63.9% vs. 57.7%. And “Overripe” signals correctly degraded to 51.4%. The scoring system distinguishes between momentum phases in a way that matters for returns.

What This Means for Retail Traders

The convergence of academic research and real-world data points to a clear conclusion:

The key distinctions that separate useful AI tools from marketing hype:

• Useful: Scanning thousands of assets for momentum patterns you would never find manually. Ranking signals by strength to focus your attention. Providing systematic entry/exit discipline.
• Hype: “AI predicts the next 10x stock.” “Machine learning beats the market every day.” “Our algorithm has a 90% win rate.” (Ask: over what horizon? With what sample size? Gross or net of costs?)

The Zarattini study, for all its caveats, gets one thing right: the value of AI is not in replacing human judgment on individual trades. It is in building and enforcing a systematic process that removes emotion from the equation and operates on time horizons where signal actually exists.

That is what our own data shows, too. Not magic. Not guaranteed returns. Just a measurable edge at horizons where the math works — and honest acknowledgment that short-term trading remains a losing game for the vast majority, AI or not.