End-to-End Deep Reinforcement Learning
Applying Deep RL to tasks such as Atari Games and Go has achieved huge success in recent years. In contrast to supervised learning where one tries to predict a value as accurate as policy, deep RL uses a policy function to accumulate as much reward as possible over time. This approach can be applied to dynamic optimization problems. For example, the optimal amount of inventory held today not only depends on demand today but demand in the future. As such deep RL is uniquely suited to solve this problem.
Traditional Data Science
Even in the age of deep learning, traditional data science has an important role in prediction problems. Models like Random Forest and Adaboost can outperform deep learning models in domains with limited data, simpler structural relationships, and larger amounts of statistical noise. This domain is still important for businesses as everything from forecasting sales to building customer profiles depend on this type of prediction problem.
In contrast to frequentist inference, Bayesian inference involves using a prior to come up with a distributional prediction for the output. For example, instead of predicting the expected amount of sales is 100 million dollars, Bayesian Inference can predict that sales are a multimodal distribution with peaks at 80 million and 120 million. As businesses are increasingly worried not only about expected sales or other outcomes but worst and best-case scenarios, Bayesian analysis can provide the tools to understand these outcomes. Unlike historical methods, state-of-the-art techniques are both fast and easy to program.
Natural Language Processing
Natural Language Processing or NLP is the use of computers to analyze language. Tools in machine learning have advanced to the point where computers can translate text at the level of humans and generate text on any topic that could potentially pass a Turing test. Businesses need to understand textual data, be it customer feedback, internal emails, or external analysis. on the world wide web. However, hiring people to read 100,000 pages of data is not feasible and one can use machines to glean human quality insights at a fraction of the cost.
As our parent organization specializes in algorithmic trading, we have expertise that only comes from being practitioners, in designing machine learning algorithms for the financial markets. Our fund was one of the firms that pioneered the use of deep Reinforcement Learning in the financial markets and have broad experience across asset classes ranging from Forex, Equities, and Commodities.