OpenAI Gym Environment for the Asset Allocation Problem
This environment implements the Asset Allocation problem.
Currently there are no transaction costs, and the assets are fixed, with data pulled from quandl.
For each assset, a random year of data is used for each episode (the year used can differ for each asset).
The portfolio is rebalanced at the start of each day. On a step, we provide action describing how we want to balance the portfolio at the start of the day. As an observation we are then given the days OHLCV for each asset, and our reward is the fractional change in portfolio value over the course of the day.
An episode terminates after 1 year of trading.
The observation at each timestep is a numpy array of shape (n, 5), containing the day's OHLCV for each of the n assets.
The action given is the relative allocation of funds between each asset. Each 0 ≤ aᵢ ≤ 1, and if Σaᵢ > 1 we normalise such that Σaᵢ = 1. If Σaᵢ < 1 the remaining funds are put into 'cash', whose value does not change with time.
Rebalancing is done at the start of a day, so assets are bought and sold at the day's open price.
The reward given is (valueₜ - valueₜ₋₁) / valueₜ₋₁, ie the change in portfolio value from the start of day t-1 to the start of day t.
- Try out a baselines agent
- Add more assets
- 'short' for each asset, with price series mimicing the result from shorting the asset?
- Add transaction costs
- Adds some complexity with rebalancing portfolio
- Add more than just
OHLCVdata- eg some fundamental stuff, technical indicators, whatever