Financial Feature Engineering – How to Research Alpha Factors

Algorithmic trading strategies are driven by signals that indicate when to buy or sell assets to generate superior returns relative to a benchmark, such as an index. The portion of an asset's return that is not explained by exposure to this benchmark is called alpha, and hence the signals that aim to produce such uncorrelated returns are also called alpha factors.

If you are already familiar with ML, you may know that feature engineering is a key ingredient for successful predictions. This is no different in trading. Investment, however, is particularly rich in decades of research into how markets work, and which features may work better than others to explain or predict price movements as a result. This chapter provides an overview as a starting point for your own search for alpha factors.

This chapter also presents key tools that facilitate computing and testing alpha...

Alpha factors in practice – from data to signals

Alpha factors are transformations of raw data that aim to predict asset price movements. They are designed to capture risks that drive asset returns. A factor may combine one or several inputs, but outputs a single value for each asset, every time the strategy evaluates the factor to obtain a signal. Trade decisions may rely on relative factor values across assets or patterns for a single asset.

The design, evaluation, and combination of alpha factors are critical steps during the research phase of the algorithmic trading strategy workflow, which is displayed in Figure 4.1:

Figure 4.1: Alpha factor research and execution workflow

This chapter focuses on the research phase; the next chapter covers the execution phase. The remainder of this book will then focus on how to leverage ML to learn new factors from data and effectively aggregate the signals from multiple alpha factors.

Alpha factors...

Building on decades of factor research

In an idealized world, risk factors should be independent of each other, yield positive risk premia, and form a complete set that spans all dimensions of risk and explains the systematic risks for assets in a given class. In practice, these requirements hold only approximately, and there are important correlations between different factors. For instance, momentum is often stronger among smaller firms (Hou, Xue, and Zhang, 2015). We will show how to derive synthetic, data-driven risk factors using unsupervised learning—in particular, principal and independent component analysis —in Chapter 13, Data-Driven Risk Factors and Asset Allocation with Unsupervised Learning.

In this section, we will review a few key factor categories prominent in financial research and trading applications, explain their economic rationale, and present metrics typically used to capture these drivers of returns.

In the next section, we will demonstrate...

Engineering alpha factors that predict returns

Based on a conceptual understanding of key factor categories, their rationale, and popular metrics, a key task is to identify new factors that may better capture the risks embodied by the return drivers laid out previously, or to find new ones. In either case, it will be important to compare the performance of innovative factors to that of known factors to identify incremental signal gains.

Key tools that facilitate the transformation of data into factors include the Python libraries for numerical computing, NumPy and pandas, as well as the Python wrapper around the specialized library for technical analysis, TA-Lib. Alternatives include the expression alphas developed in Zura Kakushadze's 2016 paper, 101 Formulaic Alphas, and implemented by the alphatools library. In addition, the Quantopian platform provides a large number of built-in factors to speed up the research process.

To apply one or more factors to an investment...

From signals to trades – Zipline for backtests

The open source library Zipline is an event-driven backtesting system. It generates market events to simulate the reactions of an algorithmic trading strategy and tracks its performance. A particularly important feature is that it provides the algorithm with historical point-in-time data that avoids look-ahead bias.

The library has been popularized by the crowd-sourced quantitative investment fund Quantopian, which uses it in production to facilitate algorithm development and live-trading.

In this section, we'll provide a brief demonstration of its basic functionality. Chapter 8, The ML4T Workflow – From Model to Strategy Backtesting, contains a more detailed introduction to prepare us for more complex use cases.

How to backtest a single-factor strategy

You can use Zipline offline in conjunction with data bundles to research and evaluate alpha factors. When using it on the Quantopian platform, you will...

Separating signal from noise with Alphalens

Quantopian has open sourced the Python Alphalens library for the performance analysis of predictive stock factors. It integrates well with the Zipline backtesting library and the portfolio performance and risk analysis library pyfolio, which we will explore in the next chapter.

Alphalens facilitates the analysis of the predictive power of alpha factors concerning the:

• Correlation of the signals with subsequent returns
• Profitability of an equal or factor-weighted portfolio based on a (subset of) the signals
• Turnover of factors to indicate the potential trading costs
• Factor performance during specific events
• Breakdowns of the preceding by sector

The analysis can be conducted using tearsheets or individual computations and plots. The tearsheets are illustrated in the online repository to save some space.

Creating forward returns and factor quantiles

To utilize Alphalens, we need...

Alpha factor resources

The research process requires designing and selecting alpha factors with respect to the predictive power of their signals. An algorithmic trading strategy will typically build on multiple alpha factors that send signals for each asset. These factors may be aggregated using an ML model to optimize how the various signals translate into decisions about the timing and sizing of individual positions, as we will see in subsequent chapters.

Alternative algorithmic trading libraries

Additional open source Python libraries for algorithmic trading and data collection include the following (see GitHub for links):

• QuantConnect is a competitor to Quantopian.
• WorldQuant offers online competition and recruits community contributors to a crowd-sourced hedge fund.
• Alpha Trading Labs offers an s high-frequency focused testing infrastructure with a business model similar to Quantopian.
• The Python Algorithmic Trading Library (PyAlgoTrade) focuses...

Summary

In this chapter, we introduced a range of alpha factors that have been used by professional investors to design and evaluate strategies for decades. We laid out how they work and illustrated some of the economic mechanisms believed to drive their performance. We did this because a solid understanding of how factors produce excess returns helps innovate new factors.

We also presented several tools that you can use to generate your own factors from various data sources and demonstrated how the Kalman filter and wavelets allow us to smoothen noisy data in the hope of retrieving a clearer signal.

Finally, we provided a glimpse of the Zipline library for the event-driven simulation of a trading algorithm, both offline and on the Quantopian online platform. You saw how to implement a simple mean reversion factor and how to combine multiple factors in a simple way to drive a basic strategy. We also looked at the Alphalens library, which permits the evaluation of the...