Statistical Arbitrage and Sentiment with NLP

What did we do?

We use the mean reversion attributes of spreads of metals as well as sentiment analysis to develop a model that outperforms its benchmark.

How did we do it?

We observed the different cointegration confidence levels of different metal spreads. We use an aggregation of spreads to acquire more stationary with which we can more confidently apply mean reversion strategies. The overall strategy combines signals from the mean reversion aspect of the spreads with sentiment analysis.

How did we implement it?

The sentiment analysis uses natural language processing with Python to locate different instruments that relate to commodities, using key words, and gets their sentiment. We weight this sentiment with each company's relevance to obtain an overall daily sentiment for commodities. These signals were combined and a strategy was backtested in Excel and VBA.

Built With

Submitted to

Algothon 2018 @ BlackRock

Created by

I worked on the Python part of the project focusing on statistical arbitrage. Specifically, I was tasked with finding out if the instruments were cointegrated by performing tests (Dickie-Fuller, etc.) and backtesting. I am also in the process of training a neural network with stock data but this was not implemented in the final project due to lack of time. I also helped in the data-mining process for sentiment data.

Tudor Trita
I worked on the Mean Reversion (Statistical Arbitrage) strategy. I calculated and normalized the spreads of the four metals, weighting each spread relative to the co-integration p-value, which was key to maximizing PnL. I created a comprehensive backtesting tool with VBA in excel, to test both momentum, mean reversion and sentiment trading models This tool supports many price feeds with multiple strategies, with clear and easy to use optimization of parameters. Inculded in the backtest are T-Cost and slippage. I worked a lot on the optimization of the parameters used in the mean reversion model, being careful not to overfit. I also backtested both strategies across different times, benchmarking against a relative index, and then combined signals from both strategies to obtain our final strategy, reducing risk and T-Cost. I also output all of the backtesting data and made the presentation.

Leon Wu
I also worked on the Natural Language Processing but using different datasets such as articles in pickle files. I tried to go through the files to get article titles and match them with relevant keywords to tell if certain company was doing good or not. I thought of creating a sort of dictionary to add more keywords that would predict the growth or stagnation of a company. However I could not implement my ideas as this was my first python project and so I ran out of time.

Hyung Jip LEE
I worked on the Natural Language Processing part of the project. learning a lot about using Python to fetch and parse HTML. I also created several helper scripts to organise the information in several CSV files before these were added to the main project.

Abel Shields

Updates

Leon Wu started this project — Oct 21, 2018 06:23 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.