The module inputs an excel file containing information like opening, closing, adjusted closing, high and low prices and the volume of trades per day. The information is used as input to the networks and the output will be the proportions of the assets. This program uses actor-critic algorithm to get the optimal trading strategy
The excel data to be used in this problem shall be built by using the dataBuilder module. Please note that each sheet in excel shall contain the data for Open, High, Low, Close, Adj Close, Volume. Also kindly ensure that the naming of these sheets as they are case sensitive and have it the same as described above.
The states to the actor and critic networks are the OHCLV data of individual stocks for the past twenty days. The actions from the actor network will be the proportion of the asset at each time step.
The environment in this problem is the daily price movements of the assets. Each time step corresponds to moving over a single day with asset proportions defined by the actor network, before the start of the time step. Here we assume that we somehow buy and sell the assets before the trading day begins.
The reward for the actions is the change in the portfolio value, which is calculated by multiplying the percent change with the asset proportions and summing them all up. Note that in this problem we assume a starting captial with currency denomination of SEK. This is because popular brokers like Avanza, Aktieinvest charges comission based on how much money is used up in each transaction order and there are different brackets for this with different comission rates. Due to this setup we shall assume that each asset shall be bough with maximum of X SEK and the proper comission rate shall be applied while calculating the portfolio value. Note that the portfolio value will be sum of the value of assets and remaining cash.
Both the actor and critic network will be made of sequentially built dense NN layers. The output from the actor network will be of shape (number_of_assets,1), whose activation will be a softmax activation. The reason is in two parts, one the sum of the output will add to 1, which is required because we don't consider shorting any stocks in this problem. Secondly the value of each entity in array will be in range [0,1], which potrays the proportion of a stock in the portfolio.
The data file contains every day price of 21 different stocks (primarily from Swedish market), from 2011 Jan to 2022 Oct. The training data will be from 2011 Jan to 2020 Dec, the test data set will be from 2021 Jan till end of the data. There is no validation set to tune the hyperparameters in this case.