Henri Waelbroeck, Portware
"Looking at value-weighted results, we find that applying deep learning in alpha profiling can improve the five-year return of a fund...enough to move a typical large-cap fund 15% in peer rankings."
Portfolio Managers (PM) seek long-term returns, but the trader needs to execute over hours or days.
Should I trade fast or slow? If the stock runs, should I chase or pull back? And vice-versa, if there is a price/size opportunity, should I take it or would I get run over by a freight train?
Machine learning techniques can help a trader answer these questions - not as a crystal ball, but in a statistical sense.
If one strategy is likely to result in a lower execution cost 60% of the times, using it can save the portfolio manager money in the same way as statistical arbitrage generates alpha.
The idea sounds simple, but to make it work requires dealing with a major problem: when looking at historical records of executed trades, observed prices are affected by the market impact from trading.
If a class of trades in a dataset seems to be associated with adverse average returns, more likely than not it is because this group of trades caused a lot of impact - perhaps because the orders were large relative to available liquidity, or volatility was unusually high. A naïve application of machine learning would "learn" that these trades develop adverse price moves and should be traded closer to arrival price.
In the example shown in the figure, the data only show realised prices (solid line). The only way to know that this was a low urgency trade would be to estimate market impact costs and draw the impact-free price (dashed, bottom). If one looked at realised returns without impact adjustments and decided to trade faster to capture the prices closer to arrival, the result would be even greater impact (dashed red line). To determine the optimal execution strategy, it's essential to adjust for market impact.
We first implemented this methodology with the Alpha Pro system in 2007, providing a classifier that selected the lowest-cost execution strategy based on order attributes such as size and PM name together with market drivers such as momentum, news and any overnight price gap.
The model automatically assigned an incoming order to an optimal strategy using a decision tree. Each node in the tree is determined using machine learning techniques to split the data into groups that share similar short-term alpha characteristics.
Alpha Vision added a user interface enabling traders to see the factors that are associated with short-term alpha, and see the execution plan as well as real-time updates on the progress of the execution. The next step in the evolution of the platform was the introduction this year of adaptive deep learning capabilities.
In conventional machine learning, coefficients are learned from past data and predictions rely on the assumption that the underlying structure has not changed. For example, models might use parameters like historical volatility and daily volume. What financial institutions need is an approach where prediction agents can access the outputs of other agents, opening the door to deep model architectures. With this type of system, each prediction spawns a timed event which feeds back into the learning process.
Deep learning uses multi-layered model architectures where the first layers produce forward estimates of the key parameters to be used in higher levels. Besides providing more accurate alpha profiling (see chart at the bottom), the adaptive deep learning paradigm leverages real-time monitoring of the performance of the prediction models to track confidence and alert users to any change in the underlying system, known as "concept drift".
The diagram shows the pre-trade system used to select an optimal algorithm.
Volume and volatility can vary greatly from day to day, especially in small cap stocks - to predict trade urgency and address "what-if" scenarios such as the appearance of a competitor or contra while executing, model accuracy can be greatly enhanced by using predicted volatility and volumes rather than historical averages as drivers.
The diagram shows some of the models used in the trade control and monitoring system. To illustrate model performance, we show out-of-sample testing results for trade urgency, defined as the expected savings in basis points from executing with a fast execution strategy vs. a slow one.
The definitions of "fast" and "slow" are dependent on trade size and tailored to the client's typical workflow - for example, for small trades the fast option might be to trade using a 10% average participation rate, whereas the slow option would be to trade over the day using a VWAP algo.
The figure shows the actual difference in implementation cost between fast and slow strategies on the y-axis in the 2013-2014 test period, as a function of the predicted urgency in a model trained on data from 2009-2012.
The vertical range of almost 20bps is 2x the savings that can be achieved over random strategy selection in the top and bottom 5% urgency cases, on a flat average basis.
Looking at value-weighted results, we find that applying deep learning in alpha profiling can improve the five-year return of a fund by 50-200bps, depending on turnover and market capitalisation - enough to move a typical large-cap fund 15% in peer rankings.