Deep Learning in Finance 2016
This past Friday the 23rd, I was in the Deep Learning in Finance summit in London.
It was a one day, single track conference with three main topics:
- Financial forecasting;
- Virtual assistants (=bots. I guess the first one sounds more friendly to the elders);
- Text analysis for risk & decision making.
All the talks were limited to twenty minutes (18 talk + 2 minutes Q&A), which is a good thing, but many were unrehearsed. As a conference organizer and speaker I know this is hard: everybody's time is limited, and if you're not a natural you'll need plenty of time to deliver a good talk. This was exacerbated by the fact that most speakers were founders, co-founders, and C-level executives of mostly small companies: almost none of them trained to publicly speak1.
As for the topics: the financial forecasting one was by far the one with the highest attention and talks: predicting the stock market is the holy grail of Finance. Having done data science in financial markets, I can say that there are two main things to consider:
- High Frequency Trading (HFT) works and it's all about performance of a well understood market with clear rules. No deep learning needed here;
- Traders do not agree on which strategy is the best; deep-learning is very good at outperforming humans in tasks where humans are good, such as image recognition, speech recognition, etc. Therefore it's difficult to train a model saying: this is good, this is bad, now go and do your magic2.
Something that was not touched upon is long term investment, i.e. buying stocks and then staying in the market for 20-30 years. This area, where you can get spectacular return, often requires profound knowledge of the companies involved and the markets they're in. You need to read documents, look at management, competitors, etc. To systematically approach it, you need to structure tons of unstructured data, feeds, make comparison, etc. This is probably an area where deep learning can be very fruitful; however I suspect the data gathering effort is much higher and the sample of companies where you can get that data is too limited for an algorithm to pick patterns up.
I'd love to be wrong though.
The second theme, bots, had several talks but they really seemed as mostly sales pitch3. If there was any theory, there were no definitive results. Remember folks: if you haven't results and you want me to invest money already, I'm acting as a VC for you, and not as a customer.
The third theme was half sale pitch, half an academic talk, but very interesting none the less.
Take home message
Deep learning in finance still hasn't penetrated the industry as in other areas. Many people are still beginning to see if they can get something out of it. The one who did and already have results can't wait to share their stories, even if they are in the very early stages (I think only Alpaca run in production right now, with real results in the forecasting area). The talks were thus spot on for most of the audience: they want to know who is one step ahead of them to be reassured they are in the right direction.
Another interesting bit caught my hear: You can't recommend an investment to a customer if you can't explain it. Deep learning has a big issue here as a neural network is not among the easiest things to explain.
I've added small companies right there as larger companies are usually more structured in the formative process of their management. ↩
You can use the future to see whether a prediction was be good or not, but why did the stock move up or down will still be debatable among traders and expert. ↩
A small pitch at the beginning or the end is fine, just don't bark all the time. ↩
Follow us for more of this
Reactive Web Service Client with JAX-WS
May 25, 2018
Fairness in Machine Learning with PyTorch
May 22, 2018
Elitist shuffle for recommendation systems
May 13, 2018
GoDataDriven open source contribution: April 2018 edition
May 06, 2018
Event Report: Blockchain Ask Me Anything
May 03, 2018
Towards fairness in ML with adversarial networks
April 27, 2018