Is Machine Learning Better For Investing Or Gambling?

Is Machine Learning Better For Investing Or Gambling?

JPMorgan’s Head of Equity Lee Spelman on Active Vs Passive Investing

Which data set theoretically and why? 

Is a machine learner better at running a casino or portfolio, why theoretically? 

In the past, it was hard to imagine how machine learning could be applied to the finance and entertainment  area. However, with the development of machine learning, machine learning has become an important  role in investing and gambling. Considering both the data set and machine learner, I believe machine  learning is more suited for gambling than investing. 

Moreover, the gambling dataset is more suitable than investing the data set to train a model.

Firstly, the Gambling game has a clear information boundary and limited states in the dataset. Let’s take the card game as an example. There are only 52 playing cards in total. And if one draws two cards initially, there are only 1326 possibilities for the initial state in total. However, for the investment, it is difficult to determine the information boundary of the whole problem in advance. Investing is an open-ended problem. One where any information may have an impact on the decision. So it becomes difficult to clearly define which features make the best inputs.

For example, in addition to historical transaction data, we can also extract valuable information from news or public opinion. However, it may contain a lot of useless or even false information.  This will produce a data set with a huge amount of information but a very low signal-to-noise ratio. With such a database, it is difficult to obtain satisfactory results during the training process.

In addition, although finance database systems, like Bloomberg and Wind, can bring some convenience for fetching transactional data, there may exist some problems like missing value or dirty data. If those problems do not become properly addressed. The performance of machine learning models will fluctuate.  

When it comes to the machine learner, it is better at running a casino rather than a portfolio, theoretically.

As a machine learner, its task is to find the mapping relationship between the input and output. As a result, is essential to defining them clearly. The gambling game has quantifiable standards, so it has a clear definition of input and output consequently. From Blackjack to Texas Hold’em poker, the fundamental purpose of gambling games is to improve the ability to win. In a machine learner, this kind of ability becomes assessed by a win or a loss according to the objective rules. During the training process, a machine learner aims to pursue a victory, which is the same as gambling’s fundamental purpose.

On the other hand, although we want to improve the investment ability of a machine learner, there is no clear or quantifiable criterion to use. First, we can only use the historical data, like yield and information ratio, as the proxy index.

The purpose of the machine learner is to improve its performance on those simple metrics, while they are not a fully accurate representation of investment capacity. So, there is a gap between the targets of machine learners and investment. The overly one-sided pursuit of optimization of these metrics also tends to fall into the risk of overfitting. Second, as the environment of investing is varying, it is difficult to determine whether this strategy is a good one. The strategy which works well on a historical data set is not guaranteed to achieve a good performance in the future.  

To sum up, machine learning is better at running a casino, because the gambling data set has a clear  information boundary, and its machine learners are well-defined.

Written by Yiwen Song

Back To News

Is Machine Learning Better For Investing Or Gambling?