DSAI HW 2 : stock_trading

Members: 陳香君、方郁文

Goal

Predict the "Action to take" in order to "Maximize" the profit you gain in the stock market.

According to hw2 description
The action should be one of these three types:
1 → means to “Buy” the stock. If you short 1 unit, you will return to 0 as the open price in the next day. If you did not have any unit, you will have 1 unit as the open price in the next day. “If you already have 1 unit, your code will be terminated due to the invalid status.“

0 → means to “NoAction”. If you have 1-unit now, hold it. If your slot is available, the status continues. If you short 1 unit, the status continues.

-1 → means to “Sell” the stock. If you hold 1 unit, your will return to 0 as the open price in the next day. If you did not have any unit, we will short 1 unit as the open price in the next day. “If you already short 1 unit, your code will be terminated due to the invalid status.“

In the final day, if you hold/short the stock, we will force your slot empty as the close price of the final day in the testing period. Finally, your account will be settled and your profit will be calculated.

How to run

Python Version: 3.6.12

Args 介紹

    '--training':
                    train file path, default training.csv

    '--testing':
                    test file path, default testing.csv
                    
     '--output':
                    output file path, default output.csv

    '--epoch':
                    number of epochs to train, default 30
                       
    '--reference':
                    number of reference days, default 100

    '--forecast':
                    number of predict days, default 10

Execution

Run the command below and will get action of each day

python main.py --training training.csv --testing testing.csv --output output.csv

Methods

Implement Convolutional Neural Network as Predictor

• Use CNN as encoder(feature extractor) and torch.nn.Linear as decoder
• Input 100 or MORE days of features (Open, High, Low, Close)
• Predict 10 or MORE days of Open price

Model Architecture

Loss

• Ground Truth(GT): nn.Softmax(open_values_of_forecast)

• Predict(PR): nn.Softmax(model.output)

• Ground Truth Raw(GTR): open_values_of_forecast

• Predict Raw(PRR): model.output

• loss: weight_a * nn.BCELoss(GT, PR) + weight_b * nn.MSELoss(GTR, PRR)

We care more about the "trend" of the stock price instead of the "absolute value", so we apply softmax to "Ground Truth" and "Predict", and use BCE loss to predict the "Trend"
However we still add MSELoss to constrain the predict value
Default weight_a and weight_b values are 1000 and 0.1 due to balancing reasons.

Evaluation

Kendall tau distance

https://en.wikipedia.org/wiki/Kendall_tau_distance

• A metric that counts the number of pairwise disagreements between two ranking lists.
• The lower, the better.
• #_of_opposite_pairs / total_pairs

for example:
x1 = [1, 2, 3, 4, 5] x2 = [6, 7, 8, 9, 10] x3 = [10, 9, 8, 7, 6]

k(x1, x2) == 0
k(x1, x3) == 1
good for predicting trend

Hyper params

Optimizer = Adam
Learning Rate = 3 * 10^-4
Epoches = 30

• Line chart of trading training.csv

training steps