We are inspired by the frontiers of machine learning research at the current moment, thus we were keen on conducting research and hypothesising the use of the latest techniques in algorithmic trading. One of these frontiers is LifeLearning or NeverEnding(1) learning (DeepMind). We were keen on applying this to the alpha-decay problem, therefore making this a general solution to many alpha-generating models (also industry and market independent). We hypothesised that alpha-decay can be due to an increase in market efficiency and also market participants becoming aware of the deployed alpha-generating models - in our model we would aim to continuously teach the model as the market develops, in the process combining and adding new alpha-generating models and letting the model decide how the markets are reacting to each model as time goes by.

Our methodology was to compare a dataset on two different models, (1) new LifeLearning deep learning model (2) and a standard Keras model and compare the alpha-decay in both models. In a proper environment, we would take a model that has suffered from alpha-decay

(1) NED

What it does

The framework extends existing deep learning algorithm to evolve itself to cope with tasks of similar but different nature. In order to share the knowledge of similarity and adapt to the changes, when the deep network is faced with a new task. It selectively retrain certain neurons in different layers in order to capture the similarity. To account for the change in task nature, new neurons are added on a ad-hoc basis to all the layers of the network. Then new neurons and selected old neurons are trained for the new tasks. Thus the deep network is able to evolve itself to cope with changing tasks.

How I built it

We based our framework on an ICLR paper published in the first half of 2018 link. We have been working with some of the published code, however, since they are not directly relevant to us we have been tinkering with the code for our purposes.

At the same time, we have been using Spark to manage and work with some of the Quandl data to prepare training and test datasets for the deep learning model.

Accomplishments that I'm proud of

Being unique and applying a new and emerging topic in Machine Learning to the real world.

What's next for Axolotl: Framework for lifelong learning trading algos

To make it a viable and tradable opportunity we would like to take into account the following:

Refine transaction cost and PnL measures

Optimize transaction costs

Risk Adjusted performance a. Sharpe Ratio b. Performance Attribution - parametric models for attribution ( e.g. curve risk )

Market Risk consideration a. Value at Risk b. Expected Shortfall


Silver DL, Yang Q, Li L. Lifelong Machine Learning Systems: Beyond Learning Algorithms. InAAAI Spring Symposium: Lifelong Machine Learning 2013 Mar 25 (Vol. 13, p. 05).

Parisi GI, Tani J, Weber C, Wermter S. Lifelong learning of human actions with deep neural network self-organization. Neural Networks. 2017 Dec 1;96:137-49.

Chen Z, Liu B. Lifelong machine learning. Synthesis Lectures on Artificial Intelligence and Machine Learning. 2016 Nov 7;10(3):1-45.

Built With

Share this project: