An alpha is a method of making predictions about future asset price changes. For example, an alpha might be a computer program that predicts future returns of a particular set of stocks.

Alphas we cover in this chapter are fully systematic and can be expressed by a concrete piece of code. The alpha will typically make predictions at some periodic frequency, for example, once per day. Its predictions will then be represented by a number for each asset it intends to predict.

A simple alpha might be, for each day, assigning a prediction of +1 to all stocks that went down yesterday and –1 to all stocks that went up. This is a valid alpha for us because it systematically generates a specific prediction for a set of assets at a specific frequency.

Alphas are predictive models, but without a way to implement those predictions, there is no way to realize the potential profits that those predictions might generate. Typically, an alpha is utilized as a component of a trading strategy, which converts the alpha’s predictions into actual trading decisions. The strategy is largely driven by the combined predictions of its alphas, but it also considers practical issues such as transaction costs and portfolio risk before actually executing a trade.

CATEGORIZATION OF ALPHAS

Alphas may be categorized into three major groups according to the types of instruments traded, such as stocks, exchange-traded funds, currencies, futures, options and bonds, etc. Alphas may trade single or multiple types of instruments. They could be developed for one specific country or multiple countries combined, or even the global market.

According to the time the alphas use the information, and the frequency at which the predictions are generated, we may categorize those alphas into the following groups:

1. Intraday alphas: rebalanced during trading hours of the day. They can also be grouped as follows:
   1. Rebalance at each interval, e.g. 1 min/5 min/15 min, etc.
   2. Rebalance triggered by some events such as ticks/orders/fills or predefined events.
2. Daily alphas: rebalance every day. These types of alphas can be broken into further subgroups by the time the information is used:
   1. Delay N: use data of N days ago.
   2. Delay 0 snapshot: use the data before a certain time snapshot.
   3. MOO/MOC: alphas trade at market open/close auction session.
3. Weekly/monthly alphas, rebalanced every week/month.

DEVELOPMENT OF AN ALPHA

An alpha is developed by using public information. The more efficient the process, the better performance the alpha can achieve. One can find alphas either by sourcing public information or building specific models to process the information. Alphas can be generated by searching signals/patterns from the informational spaces. Typical sources are as follows:

1. Price/volume. We can use technical analysis or prediction/regression models based on the price/volume.
2. Fundamentals. By analyzing the fundamentals of each company automatically, one can build fundamental alphas. Such alphas typically have very low turnover.
3. Macro data, such as gross domestic product numbers, employment rates. Such numbers have big impacts on the financial markets.
4. Text, such as Federal Open Market Committee minutes, company filings, papers, journals, news, or even information in publicly available social media. It’s necessary to quantify the text into numbers (eventually number of shares to buy/sell). Text data includes both current and future events.
5. Multimedia such as videos/audios can also be used as information sources. The techniques to process video/audio are pretty mature. For example, one can simply use Text-To-Speech techniques to extract text information from the video/audio and then build models on the text information.

Sometimes alphas are not derived from the models of information directly. This information may be used to improve the performance of alphas or generate alphas. Some examples are listed below:

1. Risk factor models: by controlling risk exposure or eliminating risk exposure to some factors, one can improve the alpha’s performance.
2. Relationship models: e.g. instruments typically correlated with each other to some extent. Some may lead or lag with others, thus they generate the opportunities for arbitrage.
3. Microstructure models to improve the execution performance of real trading.

Today, with information growing explosively, extracting signals from an ever-expanding ocean of noise is more and more challenging. The solution space is non-convex, discontinuous, and dynamic; good signals often arise where least expected. How does one extract such signals? By limiting the search space, using methods previously used by treasure hunters, such as searching in the vicinity of previous discoveries, conserving resources to avoid digging too deeply, and using validated cues to improve the probability of a find. At the same time, always allocate some processing power to test wild ideas.

VALUE OF AN ALPHA

The ultimate test of alpha value is how much risk-adjusted profit it adds to the strategy in which it is trading. However, in practice, this is difficult to measure precisely because:

• There is no canonical strategy in which an alpha may be used, and the exact strategy in which the alpha will be used may not be known at the time of alpha design.
• Even given a level of risk aversion and a specific strategy, there can be non-linear effects in the combination that make it difficult to precisely attribute profit back to individual alphas.
• All that being said, we can still:
  • Make useful predictions about whether an alpha will add value in strategies.
  • Give a reasonable estimate for how much an alpha contributed to the strategy’s profit.

PRACTICAL ALPHA EVALUATION

Since we may not know the target trading strategy ahead of time, considering an alpha on its own, how do we know if it is good or bad? Alternatively, when we make a change to an alpha, how do we know if it is an improvement? To answer these questions, we need some measurements that can help us predict if it will add value to a typical strategy.

A typical method for collecting measurements about trading strategies is to run a simulation (i.e. backtest) and measure characteristics, such as information ratio. One way we can make analogous measurements for an alpha is to do a mapping of its predictions to a trading strategy. We can then assume the predictions made by the alpha are positions that a strategy would take in the specific asset. Equivalently, the trades of the strategy would be the change in the alpha’s predictions. One issue with this method is that alphas will often not map to good strategies on their own because they are designed to predict returns, not make profitable trades. We can address this by charging zero or very small trading cost in our simulation.

Once we have constructed a simulation such as the one described above, we can now take some measurements:

• Information ratio: The mean of the alpha’s returns divided by the standard deviation of the returns. Typically, we report this as a daily quantity:
  • Roughly measures how consistently the alpha makes good predictions.
  • The information ratio combined with the length of the observation period can be used to determine how confident we are to determine that the alpha is not some random noise.
• Margin: The amount of profit made by the alpha in the simulation divided by the amount of trading that was done:
  • Roughly measures how sensitive the alpha is to transaction costs. Higher margin means that the alpha is not much affected by trading costs.
  • Alphas with low margin will certainly not work as strategies. And they won’t add value unless they are very different from the other alphas in the strategy.
• Uniqueness: Could be defined as the maximum correlation of the alpha to others in the pool of alphas:
  • Lower correlation will tend to mean that the alpha is more valuable.

More complex tests can also be developed. For example, it can be useful to test if the alpha has good information ratio on both liquid stocks (stocks with high trading volume) and illiquid stocks. If the alpha is only predictive on illiquid stocks, it may have limited usefulness in a strategy that intends to trade at very large size.

FUTURE PERFORMANCE

All of the measurements in the preceding section are intended to compare two alphas where we have no additional information other than their actual predictions. However, additional information, such as how the alpha was constructed, can yield useful information in determining whether the alpha will make good predictions going forward. Ultimately, what is important is whether the alpha makes reliable future predictions, not historical predictions.

Suppose we have an alpha with high information ratio, but it was built by taking rules with little economic explanation and optimizing the parameters of said rules to the historical data. For example, suppose the alpha had 12 parameters, one for each month (x1…x12), and suppose the alpha rule is simply to buy x1 dollars of all stocks in January, x2 dollars of all stocks in February, etc. If we optimize x1–x12 over the past year, we would get pretty good predictions for last year, but there is no reason to think they would work going into next year.

In general, each optimization or improvement made to an alpha after observing historical data will improve the alpha’s historical performance by some amount, and its future performance by some different, usually smaller, amount. Special care should be taken by the alpha designer to ensure that changes are expected to improve the alpha going forward.

When changes to the alpha yield very small (or even negative) improvements to the future predictions compared to large improvements of historical predictions, the alpha is being “overfit” to the historical data. Alpha designers can measure the effect of this overfitting by looking at the performance of their alphas on data that was not used in alpha construction (out-of-sample data) and comparing it to the data used while improving the alpha (in-sample data). Comparison of in-sample to out-of-sample performance is useful not only on the alpha level but also in aggregate across all alphas of a given designer, or on groups of alphas from a given designer. These comparisons on groups of alphas can measure the tendency of a designer’s methodology to overfit.