CORN algorithm

This repo aims to implement the CORN algorithm in Python 3. CORN stands for CORrelation-driven Nonparametric and was first introduced by Bin Li, Steven C. H. Hoi and Vivek Gopalkrishnan in 2011.

(LI, Bin; HOI, Steven C. H.; and Gopalkrishnan, Vivek. CORN: Correlation-driven Nonparametric Learning Approach for Portfolio Selection. (2011). ACM Transactions on Intelligent Systems and Technology. 2, (3),. Research Collection School Of Information Systems. Available at: http://ink.library.smu.edu.sg/sis_research/2265 )

The implementation here uses the same approach as to optimize the weights with the correlation-similiar-set but instead of purely maximizing the return, paying attention also to minimizing risk. Also this implementation differs in a way, that it does not use the correlated window, but the subsequent window (as this is the one for which the optimization is done).

The codebase provides functionality to download stock price data using Yahoo Finance API, perform portfolio optimization with the minimize function from Scipy, and benchmark the performance of the optimized portfolio against an equally-weighted pendant strategy and a simple SPY buy-and-hold strategy. The portfolio optimization problem is solved using quadratic programming and is subject to constraints on asset weights and a return target.

Installation

To use this codebase, you need to install the following Python packages:

• numpy
• pandas
• yfinance
• numba
• scipy
• matplotlib
• seaborn

Usage

1. Run the get_data function with a list of stock symbols to download and clean historical price data. The function checks that there is enough data for all symbols in the list, prunes the DataFrame to contain only the common time range of all symbols, and saves the log returns to a CSV file.

2. Run the csv_to_pd function to load the log returns into a pandas DataFrame.

3. Run the optimize_portfolio function with the log returns DataFrame to obtain optimal asset weights. You can optionally set a return target.

4. Run the plot_weights function to plot the evolution of the asset weights over time.

5. Run the benchmarking function with the optimal weights, the log returns DataFrame, and a window size to compare the performance of the optimized portfolio against a benchmark index (universal portfolio) and a simple buy-and-hold strategy (S&P 500 index). The function saves a plot of the cumulative returns to a PNG file.

Be aware!

1. If you read papers regarding CORN or any derivatives of it, you really need to pay strong attention to the assumptions. Especially, transaction costs and rebalancing period. Most of the papers assume no tx-cost and a daily rebalancing.

2. The code already makes use of numba (a JIT compiler that translates a subset of Python and NumPy code into fast machine code), to accelerate the calculations. But still, since we have to loop over the whole time series every day in order to calculate the correlation-similiar-set (correlated with the current window) it can take forever if you provide more than like 4-5 years of daily data. It can be optimized for performance even more, but I haven't had the time.

Performance

300 Trading days

600 Trading days