AI-based stock trading, a record-breaking competition on Kaggle and more stories cherry-picked from all the interesting ML- and AI-related news from September. Right here in the AI Monthly Digest.
The Digest gathers machine learning and AI news to spot the most important and interesting events and developments of the past month. The five events below were curated from last month’s events and chosen by Arkadiusz Nowaczyński and Konrad Budek from deepsense.ai team.
Deep learning takes a deep dive into the stock market
Deep reinforcement learning can be applied as a complete AI solution for algorithmic trading.
The authors of “Deep Reinforcement Learning in Portfolio Management” set out to determine whether methods derived primarily for playing Atari games and continuous control would work on the stock market. The algorithm they used, called deep deterministic policy gradient (DDPG), returned promising results in an offline backtest.
The second paper, “Deep Reinforcement Learning in High Frequency Trading,” provides convincing arguments about why AI stock trading is suitable for trading in a timescale below 1 second (High Frequency Trading). The authors did a solid evaluation of their approach with a few noteworthy tips:
- Online learning at test time makes it possible to maintain high accuracy over time;
- A small neural network is enough for this problem, meaning AI for trading can be developed on laptops;
- Predicting the next 100 ticks from the last 500 ticks works best for them.
Progress remains to be made and questions to be answered. Does this algorithm work when deployed on the real market? How much money can you actually make with it? The lack of answers is certainly intriguing, as is the fact that algorithmic trading may soon be powered mostly by Deep RL, if it’s not already. We think that the potential financial reward will push people to develop further breakthroughs in AI. After all, setting high scores in Atari games isn’t as satisfying as having supersmart AI earning you gobs of money.
For those of you who are new to the topic we recommend a good nice-to-read introduction.
A record-breaking Kaggle competition
Over 8500 data scientists on no fewer than 7000 teams took part in the Kaggle Home Credit Default Risk evaluation record-breaking competition. The goal of the competition was to predict the risk of giving a loan to a particular customer. The teams were provided with rich datasets containing historical and transactional data on the customer’s behavior.
Perfectly designed, the competition attracted teams from far and wide, mostly thanks to the outstanding dataset. It allowed the teams to harvest insights and play with data in often surprising ways. Looking to tune up their models and further polish their skills, participants engaged in discussions and peer-reviews long after the competition had ended.
deepsense.ai took part in the competition, with Paweł Godula leading a team that took 5th place overall and finished first on the public leaderboard.
Volvo trucks introduce Vera, the cabless truck
According to PwC data, by 2030 the transport sector will require 138 million fewer cars in Europe and the US, mostly thanks to the rise of autonomous vehicles and the development of new business models. What’s more, it is predicted that by 2030 autonomous vehicles will be driving 40% of all miles driven.
As a proof of concept, Volvo has brought out Vera, the cabless truck to be used in short-haul transportation at logistics centres or ports. With the fleet of vehicles able to communicate and be supervised by a cloud-based management system, the truck is an interesting glimpse of the driverless future.
DARPA announced $2 billion investment in AI
At it’s 60th anniversary conference, the DARPA (Defense Advanced Research Projects Agency) announced that it is going to invest $2 billion in artificial intelligence. The agency is known for developing cutting-edge technology, be it ARPANET, which later evolved into the Internet, or the Aspen Movie Map, which was among the predecessors of Google Street View.
According to John Everrett (via CNNMoney), the deputy director of DARPA’s Information Innovation Office, the agency’s investment is intended to accelerate the development of AI from 20 years down to five years.
DARPA’s investment is not the first a government has made in AI. The most notable example comes from the United Arab Emirates, which has appointed an AI minister.
NIPS conference sold out in less than 13 minutes
NIPS, hosted in Montreal, Canada, is currently the most important machine learning and AI research conference in the world. Initially held as an interdisciplinary meeting of experts interested in sharing their knowledge on neural networks, it has evolved into the machine learning meeting with thousands of papers sent for review. It is also a place to run competitions with the “Learning to run” in 2017 as an example.
In 2017, the tickets sold out in two weeks, a relative eternity compared to the rock concert-like 12 minutes and 38 seconds they flew out in this year. Tickets for last year’s Comic-Con, one of the world’s most beloved pop culture events, sold out in a bit more than an hour.
So, when it comes to selling tickets, Marvel superheroes would appear to have nothing on machine learning. This year’s NIPS conference will feature Henryk Michalewski, visiting professor at Oxford University and a researcher at deepsense.ai, as a co-author of “Reinforcement Learning of Theorem Proving” paper.
September has clearly shown that AI is one of the most dominant trends in modern tech. Selling out venues faster than pop culture events goes a long way to proving that a scientific conference, or at least this one, can be as exciting as a concert or show – so long as it’s about Artificial Intelligence.