Academic Signal #7

The Prediction Trap: Modern Investing Needs Classification and Causation

1. Instead of calculating expected returns, ask "What's the probability this stock becomes a winner or loser?"

Researchers at Universitat Pompeu Fabra just flipped portfolio construction on its head. Instead of asking "What will this stock return?" they asked "What's the probability this becomes a winner or loser?"

Traditional approach: Stock A has 12% expected return, Stock B has 10%. Buy A, sell B.

New approach reveals:

• Stock A: 12% expected return, but only 8% chance of top-decile finish

• Stock B: 10% expected return, but 22% chance of top-decile finish

Suddenly the "inferior" stock looks attractive.

Think of it like weather forecasting: Meteorologists are better at predicting "chance of rain" than exact rainfall amounts. Similarly, momentum and value metrics are better at identifying potential home runs vs strikeouts than forecasting precise returns.

What Actually Predicts Winners and Losers

Analyzing 7,500 stocks over 65 years, the researchers found the most powerful predictors of extreme outcomes:

The Winner Formula:

• Recent 1-month losers (short-term reversal is the #1 predictor)

• With medium-term momentum (strong 2-12 month performance)

• Plus value characteristics (high book-to-market, sales-to-price ratios)

• In sectors with positive momentum

The Loser Warning Signs:

• Recent 1-month winners hitting daily return highs

• High asset growth (empire building effect)

• Poor medium-term momentum

Testing their classification approach vs traditional expected return methods:

Equal-weighted portfolios:

• Traditional ML: 3.61% monthly returns, 2.56 Sharpe

• New classification: 6.12% monthly returns, 2.92 Sharpe

After transaction costs (where most academic strategies die):

• Classification approach: 0.84 Sharpe

• Traditional methods: 0.48 Sharpe

• Market benchmark: 0.46 Sharpe

That's 80% higher risk-adjusted returns than the market, even after realistic trading costs.

Hit “reply” and let us know if you’d like us to dig deeper into this paper or have specific questions.

2. The failure of AI in investing isn't a technology problem – it's a methodology problem

Prof. Marcos López de Prado from the Abu Dhabi Investment Authority posited at the World AI Conference that AI is great for learning how to drive a car, but fails at investing.

Driving is essentially a prediction and reaction problem: see a red light, stop; see a pedestrian, brake; follow traffic patterns. It's complex but follows learnable rules that AI excels at. Most AI systems are built for these kinds of predictive tasks – identify patterns, forecast outcomes, optimize responses.

But successful investing isn't primarily about predicting what happens next. It's about understanding why things happen and what that means for risk and reward. This requires causal reasoning, not just pattern recognition – a fundamentally different cognitive challenge that current AI struggles with.

The Core Misunderstanding

Most AI investment approaches treat markets like weather forecasting: predict what happens next, then bet accordingly. López de Prado argues that this misses the point entirely.

Investing requires causal attribution, not just accurate forecasts.

Consider his ice cream example: AI can perfectly predict that ice cream sales correlate with drownings. But an investor needs to understand the causal mechanism (hot weather drives both) to avoid the catastrophic mistake of shorting ice cream companies to prevent drownings.

Why Causal Understanding Matters

In markets, you need to know why a factor earns returns, what risks cause those returns, and how those risks interact. Without this causal understanding, you're exposed to unrewarded risks that can destroy your strategy when conditions change.

A concrete example: The value factor (buying cheap stocks) has historically outperformed. But why?

• Hypothesis 1: Cheap stocks are distressed companies that recover

• Hypothesis 2: Investors systematically overpay for growth, creating value opportunities

• Hypothesis 3: Value stocks are fundamentally riskier and earn a risk premium

Each explanation suggests different conditions under which value investing should work or fail. If you believe Hypothesis 1, you'd worry when credit markets tighten. If you believe Hypothesis 2, you'd watch for changes in investor behavior. If you believe Hypothesis 3, you'd focus on risk factor exposures.

Without understanding the causal mechanism, you can't know:

• Whether the factor will persist through regime changes

• What risks you're actually taking

• How to adjust when performance falters

• Whether poor performance means the factor is broken or just temporarily out of favor

This is why many quantitative strategies that look bulletproof in backtests fail spectacularly in live trading – they captured correlations without understanding causations.

The Right Way Forward

López de Prado proposes three critical questions for any AI investment solution:

1. Have you controlled for backtest overfitting?

• Train-set overfitting: Run sensitivity analysis

• Test-set overfitting: Use multiple-testing adjustments (Deflated Sharpe Ratio)

• Out-of-sample evidence: Implement proper embargo periods

2. What's the causal mechanism?

• Identify the risk premia attribution

• Build counterfactual reasoning capabilities

• Prepare for black-swan events with coherent stress testing

3. Is this an ensemble method?

• Avoid single points of failure

• Use bagging, boosting, stacking, and meta-labeling techniques

The Investment Assembly Line

Rather than ad-hoc AI applications, López de Prado advocates for a systematic "assembly line" approach:

1. Data Curation: Structure datasets properly

2. Feature Analysis: Test historical relationships rigorously

3. Strategy Development: Understand WHY relationships should persist

4. Testing: Calculate true false positive probabilities

5. Deployment: Optimize implementation efficiency

6. Oversight: Monitor ongoing performance vs. expectations

As López de Prado puts it: investing requires a new type of AI. One that prioritizes understanding over prediction, causation over correlation, and robustness over optimization.

Hit “reply” and let us know if you’d like us to dig deeper into this paper or have specific questions.