Mean Variance Optimization by Gradient Descent (PyTorch)

Overview

• Mean-Variance Optimization to maximize Sharpe ratio using Deep Learning (PyTorch)
  • 1 layer GRU / Transformer / TCN
  • 1 FC layer
  • loss_fn : minimize negative sharpe (or Risk Parity)
  • optimizer : SAM (base SGD)

DL Model

• GRU
• Transformer
• TCN

MVO

• Mean-Variance Optimization
• Maximize Annualized Sharpe Ratio

Optimizer

• SAM optimizer (base optimizer : SGD with momentum 0.9) was used
  • the original source is here
  • Of course, you can use Adam as base optimizer of SAM, or just Adam not SAM
    • but SAM optim (SGD) shows better performance than other options, empirically
• Learning Rate : 5e-3 (No Scheduler)

Loss Function

def max_sharpe(y_return, weights):
    weights = torch.unsqueeze(weights, 1) 
    meanReturn = torch.unsqueeze(torch.mean(y_return, axis=1), 2)  
    covmat = torch.Tensor([np.cov(batch.cpu().T, ddof=0) for batch in y_return]).to('cuda')
    portReturn = torch.matmul(weights, meanReturn)  
    portVol = torch.matmul(weights, torch.matmul(covmat, torch.transpose(weights, 2, 1)))
    objective = ((portReturn * 12 - 0.02) / (torch.sqrt(portVol * 12)))
    return -objective.mean()

Constraint

• You can configure upper/lower bound for portfolio weights
  • this bounds are handled in UB and LB key in train_config.json
  • if you don't need any bounds, just set LB=0 and UB=1
• portfolio weights are adjusted by the function below, before backpropagation

def rebalance(self, weight, lb, ub):
    old = weight
    weight_clamped = torch.clamp(old, lb, ub)
    while True:
        leftover = (old - weight_clamped).sum().item()
        nominees = weight_clamped[torch.where(weight_clamped != ub)[0]]
        gift = leftover * (nominees / nominees.sum())
        weight_clamped[torch.where(weight_clamped != ub)[0]] += gift
        old = weight_clamped
        if len(torch.where(weight_clamped > ub)[0]) == 0:
            break
        else:
            weight_clamped = torch.clamp(old, lb, ub)
    return weight_clamped

Data

• As of December 27, 2021, stocks with more than 5,000 daily price data were selected.
  • AAPL, ABT, AMZN, CSCO, JPM, etc.
• Survivorship Bias (Look-ahead Bias)
  • We didn't know in the past that these selected stocks would be in S&P500 until November 2021.
  • So, the performance might (must) be different in real stock market
• Make Dataset for Training model

python dataload/data_download.py
python dataload/make_dataset.py

Configuration

• train_config.json

{
  "MODEL" : "GRU",
  "BATCH": 32,
  "SEED": 42,
  "EPOCHS" : 500,
  "EARLY_STOP" : 50,
  "LR" : 0.005,
  "MOMENTUM": 0.9,
  "USE_CUDA" : true,
  "N_LAYER": 1,
  "HIDDEN_DIM": 128,
  "N_HEAD" : 10,
  "N_FEAT": 50,
  "DROPOUT": 0.3,
  "BIDIRECTIONAL": false,
  "LB": 0,
  "UB": 0.2
}

• data_config.json

{
  "START" : "2001-01-01",
  "END" : "2021-12-27",
  "N_STOCK" : 50,
  "LEN_DATA" : 5000,
  "TRAIN_LEN" : 63,
  "PRED_LEN" : 21,
  "TRAIN_RATIO": 0.75
}

Result

• Test Date
  • From 2017-04-11 To 2021-11-11
• Model
  • GRU (hidden_dim = 128), Dropout (0.3), Lower/Upper Bounds (0, 0.2)
• Performance (Transaction costs are NOT considered)
  • Expected Return : 0.310951 (snp500 : 0.134716)
  • Volatility : 0.180532 (snp500 : 0.166945)
  • Sharpe Ratio : 1.722411 (snp500 : 0.806953)
  • MDD : -0.179543 (snp500 : -0.233713)
• You can see the cumulative return plot in result folder

Requirements

numpy==1.20.0
pandas==1.3.4
torch==1.7.1+cu110
yfinance==0.1.67