WhiteRock (Saurav, Sim, Batuhan, Desmond)

Inspiration

The talk on human behaviour at Blackrock made us think quite closely about whether human behaviour after an event could be accurately modelled. The hypothesis was that when major events occur that contain a difference in sentiment before and after the occurrence of the event, there is a short period where trading is governed by response (algorithmic or human). Our goal was to understand the behaviour of post event trading across stocks.

The risk management talk made us look much deeper into portfolio optimisation techniques and we stumbled across hierarchal risk parity optimisation. By creating a tree and clustering stocks hierarchically based on their correlations, you effectively create a constraint on the maximum weight allocated to an asset class, sector, industry and so on. This would theoretically perform better than the standard mean-deviation optimisation by Markowitz.

We also ran other portfolio optimisation models (MVP, IVP) although testing still remains to be done.

What it does

We have created an algorithm that uses the sentiment change over a stock and fits a classifier to this to create alpha signals that define the percentage allocation of funds as well as buy/sell.

How we built it

We used python and python libraries to do all the data parsing, analysis, prediction and testing. The quandl api was used to fetch data.

Challenges we ran into

The datasets are quite big and incomplete so fetching and parsing through the data is a bit of a challenge.

Accomplishments that we're proud of

We managed to obtain a directional prediction accuracy of 53.4% and the trades executed on this signal had a very low drawdown (<6%) and a high return percentage over the S&P.

What we learned

It's very difficult to do lots of things in a short amount of time. It is difficult to find a good idea to generate alpha. Converting a signal to a strategy is also a major challenge.