In this post, I’m going to give you a brief introduction about the work we have been doing here at Predic.to. Hopefully, this is the start of a series of posts that some people might find interesting.
In the last few years, Hedge Funds and Financial Institutions have been investing in building strong Data Science and ML/DL teams. It’s no secret in those circles that there is value in using Big and Deep Data to get valuable insights about investment decisions, short term or long term.
Predic.to started as an experiment to study news, stocks and cryptocurrencies. Trying to study how news and market volatility affect pricing in the short or long term. It’s an ongoing research and everything you see here or in our website is experimental.
The goal of Predic.to is to provide people access to Deep Learning for Investing while being completely honest about risks involved in such systems.
Markets tend to overreact or underreact sometimes. We all see it. Some breaking news appear, a big move is happening. Just to slowly readjust to a new normal level again in the next few minutes, hours or days.
Although Efficient Market Hypothesis and Eugene Fama believe that stock prices always reflect all available information, there are other people that claim that there are other human behavioral factors that affect prices. I’ll redirect you to Robert Schiller and his excellent books on the topic. It’s no coincidence that both Fama and Shiller received the Nobel Prize in Economics in 2013 even though their views look different. Maybe the truth is somewhere in between? Well, we at Predic.to lean a tiny bit more towards the Shiller perspective.
Over the last year, we have been focusing on developing a number of Deep Learning and Statistical models that try to perform short-term forecasting for stock price and volatility. Our focus is around top Nasdaq companies.
Of course, to do this, we need access to lots of data. And reliable data. Data is King, and how you decide to use your Data is the Queen. Good data are also expensive. Remember that. But good data can also be mined, with dedication, in order to build a good history horizon.
So back to our story.
Those are the main fields of focus for us, and should be for anyone that is serious about Deep Learning for investing:
- Get good data with good history horizon.
- Ask the right questions — formulate the problem well and be creative.
- Use latest advances in Deep Learning (from any field) and combine models and architectures until you get somewhere.
- Test/Backtest/Validate and more importantly please be honest with yourself and your users.
- Keep trying new inputs to your models. Market conditions change frequently and so should your models.
I’m going to present some charts from our work and more insights will be published in the near future with case studies.
Below are a couple of examples of how our training/validation/forecasting pipeline looks for a few stocks. In this case forecasting horizon is 15 days ahead. The final forecast is a weighted combination of several models.
Below some examples of recent short-term forecasts generated.
Disclaimer: In all the above examples, you can see some good cases, but there are of course cases that forecast completely misses the actual price movement. That’s why it is very important to capture expected volatility and uncertainty in deep learning models by providing some accurate (as much as possible) confidence intervals. This is another work in progress for us.
If you want to know more about our work, feel free to have a look at Predic.to. We certainly have a passion for this field and we want to share some of our findings and struggles as we go on.
More posts will follow with some insights on how news appear to affect stock pricing in the short term, before re-adjusting to a new equilibrium.
In the meantime, remember that when it comes to investing, what Deep Learning can offer is identifying complex patterns that are not visible to the human eye. Whether the patterns found are useful depends on the data we feed our models, so our job is to feed quality data and make sure to generalize as much as needed but no more than that.
Deep Learning for investing can’t predict the future, can’t predict unpredictable events, but it might be able to identify complex patterns that might be useful in some way.
With a certain level of uncertainty.
Identifying and evaluating those opportunities is our goal.
