How the market will move tomorrow?

The short answer is: nobody knows not even artificial intelligence.

Lot of people have tried to predict the future and it does not take a lot before you stumble on somebody on YouTube or Google that ensures you to double your money each year because they know what the market will do the next day. Thinking a bit in perspective, let’s say that some money manager would have found this holy grail, it would not take more than 10 years for a fund to go from 100M$ to 102T$. Wait a bit longer and the fund will take over all the money in the market. Unlikely, right?

Nowadays, Artificial Intelligence is a quite popular field that is gaining traction in everything that surrounds us. Why then not to use AI to make predictions on how the market will move the next day? Let’s try and see what we get.

To do the job, I believe that the most suited AI architecture is the long short-term memory (LSTM). LSTM networks are well-suited to classifying, processing and making predictions based on time series data. It has the ability to encapsulate the notion of forgetting part of its previously stored knowledge, as well as to add part of the new information. Personally I found that the highest level of accuracy is achieved when LSTM algorithm remembers only what happened in the previous 5 days.

LSTM was applied to predict the next day price of the XIV (short selling VXX). To be honest when the simulation was completed, I thought I discovered the holy grail of investing. Simply look at the image below, the two curves are matching.  

​My enthusiasm was short lived because looking closely at the two lines, there was a gap. The gap might not seem much but if we are thinking whether to short the VXX or being in cash based on this predicted value then this difference will play a major role.

I then ran a backtest in which the strategy was in cash if the expected next day return was negative and short VXX if positive.
​
What happened? It did not go well: poor performances and lot of drawdown. This means that the gap in value between the estimation and the actual value even if in appearance it looks small then it is not as negligible as one might think.

There are some techniques that can improve the results. I’ll illustrate them in the next post but in either cases…
​
The bottom line is: tomorrow can not be predicted with high enough accuracy, neither by humans nor by machines. Which in perspective is quite nice because otherwise life would be totally boring and dull.