day, with more or less success.
CHALLENGES
Even if one can make an argument in favor of the existence of alphas under various stylized assumptions, the details of prediction in the real world are messy. A prediction with low accuracy, or a prediction that estimates a weak price change, may not be interesting from a practitioner’s perspective. The markets are an aggregate of people’s intentions, affected by changing technology, macro-economic reality, regulations, and wealth – which makes the business of prediction more challenging than meets the eye. Thus, to model the markets, one needs a strong understanding of the exogenous variables that affect the prices of financial instruments.
THE LIFE-CYCLE OF ALPHAS
A fundamental law of the markets is that any potentially profitable strategy attracts attention and attracts capital. Since the markets are a finite size, when more capital chases a strategy or employs a particular alpha, this implies that the fixed-sized pie that constituted the original opportunity needs to be sliced into multiple thinner slices. The end result is that, while alphas are born from the interaction of market participants, when they are (1) strong enough, (2) old enough, and (3) consistent enough to be statistically validated and provide the basis for profitable trading strategies, they will begin to attract capital. This capital flow will ensure that the alpha will shrink and become more volatile, until there is so much capital chasing the idea that it will stop working. However, this process will affect the markets in ways that create other patterns, perpetuating the cycle of birth and death of alphas.
DATA INPUT
In order to predict the price movement of financial instruments, alphas need data. This data can be the prices themselves or a historical record of those prices. Most of the time, however, it helps to have more information than just the prices. For example, how many shares of a stock were traded, its volume, etc., can complement the historical price–time series.
A simple diagram to represent what an alpha is doing is as follows:
Note that data quality can have a large effect on the output of an alpha. So it’s important to evaluate data quality before it is used and address shortcomings then. Issues that may affect data quality can be technical, e.g. hardware problems, or related to human error, e.g. unexpected data format change, extra digits, etc.
PREDICTIVE OUTPUT
An alpha model’s output is typically a prediction. In many markets, it’s easier to predict the relative price of a financial instrument than it is to predict the absolute price of a financial instrument. Thus, in stocks, many alpha models predict the movement of the prices of various stocks relative to other similar stocks.
Typically, alphas are implemented using a programming language like C++, Python, or any other flexible and modern language. In larger organizations, a software environment developed in-house can abstract the alpha developer from many book-keeping and data management issues, letting the developer focus on creative research and modeling.
EVALUATION
What is a good alpha? What is a bad one? There is no single metric that will answer that question. In addition, the answer depends in part on how the alpha is going to be used. Certain investment strategies require very strong predictors; others benefit, marginally, from weak ones. Some pointers to alpha evaluation are:
● Good in-sample performance doesn’t guarantee good out-of-sample performance.
● Just like in academic statistics, outliers can ruin a model and lead to erroneous predictions.
It takes a lot of in-sample and out-of-sample testing to validate an idea. The more data one has, the more confidence one can have in an alpha. Conversely, the longer the period one considers, the more likely that the alpha will exhibit signs of decay and the more likely fundamental market changes will make the alpha unusable in the future. Thus, there is a natural tension between developing confidence in an alpha and its usefulness. One must strike the right balance.
LOOKING BACK
When developing alphas, one has the opportunity to look back in time and evaluate how certain predictive models would have performed historically. And, while evaluating an alpha with backtesting is invaluable (providing a window into both the markets and how the alpha would have performed), there are a few important points to remember:
● History never repeats itself exactly, ever. So while an alpha idea may look great on paper, there’s no guarantee it will continue to work in the future. This is because of the perverse power of computation, and the ability of creative modelers to miss the forest for the trees. With computational resources, one can evaluate a very large number of ideas and permutations of those ideas. But without the discipline to keep track of what ideas were tried, and without taking that into account when evaluating the likelihood of a model being a true model versus a statistical artifact only, one will end up mistaking lumps of coal for gold.
● It’s easy to look back at history and imagine that the market was easier to trade than it was in reality. That’s because of several effects. First, you didn’t have the insight back then that you have now. In addition, you didn’t have the data back then that you have now. Finally, you didn’t have the computational power and technology to do what you can do now. Ideas that seem simple and were programmed in spreadsheets in the 1990s were actually not so simple back then, especially when one considers the research it took to get there. Every decade has its market and its own unique market opportunities. It’s up to the alpha developer to find market opportunities with the data and technology available. Looking back at previous eras, it is wrong to believe it would have been easy to predict the market and make money.
STATISTICS! = STATISTICAL ARBITRAGE
The term statistical arbitrage is another term used to describe quantitative investing. The term conjures a couple of ideas: (1) that the predictions used don’t constitute a pure and risk-free arbitrage, and (2) that statistical models are used to predict prices. While (1) is absolutely correct, (2) requires some explanation. While it is the case that certain models from academic statistics – such as time series analysis and machine learning, or regression and optimization – can be used as part of alpha development, it’s important to realize that most techniques from academia aren’t really aiming to solve the problem that is of interest to a quantitative investment firm (i.e., generating a cash flow and managing the risk of that cash flow). Thus, while looking at mean-square errors when evaluating models has some merit, it is only indirectly related to making money.
TO SUM IT UP
The existence of alphas in the market is a result of the imperfect flow of information among market participants with competing objectives. An investor with a long-term horizon may not be concerned with short-term variations in price. Conversely, a trader with a short-term horizon doesn’t need to understand the fundamental factors affecting price movements, and is rewarded for understanding short-term supply-and-demand dynamics instead. Yet these different investors coexist and arguably provide some value to the other. In addition, each investor’s actions in the market coalesce to produce patterns. The goal of an alpha researcher is to find which of these patterns are relevant to predicting future prices, subject to various constraints, like the availability of data and the researcher’s modeling toolkit. While the nature of the alphas changes with time, it is the researcher’s challenge to find which ideas are relevant today, and to implement them in a way that is efficient, robust, and elegant.
3
Cutting
