Machine learning algorithms have made their presence felt in many areas outside finance: voice recognition (Siri and Alexa), image recognition (self-driving cars), and AlphaZero (effectively the world chess champion).

While machine learning techniques have captured the interest of many investment professionals, the question remains whether they can become practical investment tools.

The term “machine learning” describes the process by which algorithms uncover relationships without explicit programming instructions. Machine learning algorithms excel at uncovering subtle, contextual, and non-linear relationships; “overfitting” can be a problem. A model picking up noise instead of signals is known as overfitting.

The authors aim to reduce overfitting and establish more reliable and useful machine learning techniques in an asset pricing context.

The authors examine two ways to reduce the risk of overfitting: (1) combining forecasts and (2) feature engineering.

Forecast combinations aim to improve on a widely used process known as model average, in which a large number of weak predictions are combined to create a stronger prediction. The authors believe the forecast is stronger when they combine forecasts from different types of algorithms and train them on smaller sets of data, or training sets.

If a sizeable number of algorithms focused on a number of training sets all uncover the same relationships and reach similar conclusions, the forecast should avoid overfitting and be more reliable.

Feature engineering is about narrowing down the questions and using only relevant algorithms to answer them. It is an effective way to reduce overfitting, the authors argue, because it enables forecasters to increase the signal-to-noise ratio even before training the algorithms.

With feature engineering, forecasters can instead predict the behavior of categories of assets. Categories may include outperformers versus underperformers, which produces less “noisy” results than trying to forecast stock price returns. These categories can be parsed into sectors and geographies to further reduce noise.

The authors’ proposed approach allows practitioners to model a real-world environment in addition to helping them avoid overfitting. Forecasters can, for instance, use training sets that are recent, exist under similar macroeconomic conditions, and occur at the same time of the year to allow for seasonal effects.

Feature engineering increases signals relative to noise by better framing questions and transforming data into patterns. Meanwhile, forecast combinations are more reliable in that they reveal patterns that are recognized by a number of different machine learning algorithms and training sets.

The authors present a case study to show that machine learning algorithms can assess a large number of company characteristics and find useful patterns while reducing overfitting. The case study sample covered small-, medium-, and large-cap stocks in 22 developed markets over a 22-year period. From this sample of stocks, the authors create portfolios encompassing 194 factors, or company characteristics, covering the range of known investment factors. They find that alpha is produced across all algorithms and training sets but is particularly strong for portfolios created using forecast combinations.

The authors compare their findings with those of a conventional linear regression approach to forecasting stock returns. Performance is positive, but alphas are considerably larger using the machine learning techniques. Tasking the algorithm with finding the 10 factors with the highest Sharpe ratios reveals that machine learning strategies are even more attractive on a risk-adjusted basis.

Applying estimated transaction costs of 15 bps to each portfolio, the authors find that alphas are still significant.

A big potential benefit of machine learning is that algorithms can discern the changing nature of relationships between factors. The case study shows that at least some of the machine learning alpha is derived from factor timing.

Machine learning can uncover complex, non-linear patterns that would be difficult to find using traditional statistical techniques.

By combining predictions from different types of algorithms and using only data and factors that are likely to provide valuable information, it is possible to amplify the signal and turn down the noise, thus improving the quality of predictions.