“If you’re so smart, why aren’t you rich?”

Let’s be very clear. It is a great thing to be smart. Being smart is not in any way an impediment to being a good investor. But just as the gift of strength can exacerbate any tendency towards being a bully, intelligence misapplied can lead a person into mistakes that a person with more limited intellect couldn’t begin to make.

We launched our Kaggle competition a few weeks ago. From their questions and ideas, it is clear that the entrants are all very smart. Which gives them the opportunity to make a classic smart person mistake: using a complex model.

There’s a paradox in quantitative finance: the fancier the model, the worse it often performs once you take it out of the lab and into the real world. Simpler approaches, though seemingly crude, tend to survive better out-of-sample. In this post, I’ll unpack why that is and show a toy example that illustrates the phenomenon.

Financial data is not like physics data. The latter follows stable laws of nature; the former is a messy product of human behavior, regime changes, and randomness. The signal-to-noise ratio is terrible. When you build a highly flexible model—for example, a deep neural net or a polynomial regression—it has the capacity to explain almost every squiggle in your historical dataset. The trouble is that most of those squiggles were just noise.

Out-of-sample, when the noise shifts, those models collapse.

Why Simple Models Generalize Better

1. Fewer Degrees of Freedom: A simple model can’t chase every twitch in the data. It is forced to focus on the strongest, most persistent patterns.
2. Bias–Variance Tradeoff: Complex models reduce bias (fit the past better) but increase variance (unstable forecasts). Simpler models have higher bias but lower variance, and lower variance usually wins in non-stationary financial markets.
3. Stationarity Assumptions: The more parameters you estimate, the more you implicitly assume the world is stable. But markets aren’t stable. Simple models depend on fewer assumptions, so fewer things can break.
4. Interpretability: You can usually explain what a simple model is doing. That makes it easier to check whether it captures something plausible (e.g. momentum) or is just an artifact.

Let’s build a simple toy model to demonstrate.

Step One:

We simulate 500 days of daily returns from an AR(1) process with a weak momentum effect:

• True expected return today depends 5% on yesterday’s return. So if yesterday’s return was 1%, then todays expected return is 0.05%.
• Annualized volatility is 30%. This is about 2% a day, so noise is much higher than the signal due to auto-correlation.

Step Two: Competing Models

• Simple model: Linear regression on yesterday’s return.
• Complex model: 5th-degree polynomial regression on the last five days’ returns.

Step Three: Procedure

• Fit both models on the first 250 days.
• Forecast returns for the next 250 days.
• Compare mean squared error and realized correlation with the true signal.

Step 4: Results (illustrative)

• In-sample (first 250 days):
  • Simple model R² ≈ 0.01
  • Complex model R² ≈ 0.35

• Out-of-sample (next 250 days):
  • Simple model R² ≈ 0.02 (small but positive)
  • Complex model R² ≈ –0.20 (completely fails, negative predictive power)

The complex model looked brilliant in-sample. It explained a third of the variance! But that was an illusion—it fit noise. The simple model looked unimpressive in-sample, but it captured the one real thing in the process: a weak persistence effect. Out-of-sample, the simple model held up, the complex one collapsed.

Lessons for Real Forecasting

1. Don’t confuse fit with edge. A great backtest means nothing if the model is just curve-fitting noise.
2. Prefer robustness to precision. A model that captures a coarse but stable effect will beat one that nails a dataset-specific pattern.
3. Validate in time, not just cross-section. Use walk-forward testing. If a strategy only works in one slice of history, it isn’t real.
4. Anchor in intuition. If you can’t explain why a model should work in economic or behavioral terms, assume it’s fragile.

Simple models are not a concession to mediocrity; they are an active defense against the treachery of financial data. They fail less spectacularly because they assume less, overfit less, and latch onto only the strongest signals. In a world where regimes shift and noise dominates, robustness beats brilliance.

In finance, as in the real world, a hammer will smash a scalpel.

Disclaimer

This document does not constitute advice or a recommendation or offer to sell or a solicitation to deal in any security or financial product. It is provided for information purposes only and on the understanding that the recipient has sufficient knowledge and experience to be able to understand and make their own evaluation of the proposals and services described herein, any risks associated therewith and any related legal, tax, accounting, or other material considerations. To the extent that the reader has any questions regarding the applicability of any specific issue discussed above to their specific portfolio or situation, prospective investors are encouraged to contact HTAA or consult with the professional advisor of their choosing.

Except where otherwise indicated, the information contained in this article is based on matters as they exist as of the date of preparation of such material and not as of the date of distribution of any future date. Recipients should not rely on this material in making any future investment decision.