justanitin/card-nse-stock-predictor

CARD NSE Stock Price Predictor

Based on: CARD: Channel Aligned Robust Blend Transformer for Time Series Forecasting (ICLR 2024)

Model Description

This model uses the CARD (Channel Aligned Robust Blend) Transformer architecture to predict NSE (National Stock Exchange of India) stock prices. Key innovations from the paper:

• Dual Attention: Channel-aligned attention (cross-stock) + Temporal attention (over time) with EMA smoothing
• Token Blend: Multi-scale token representations via head-merging
• Signal Decay Loss: Weighted MSE that prioritizes near-future accuracy (l^{-1} weighting for stocks)
• RevIN: Instance normalization for handling non-stationary financial data
• Dynamic Projection: Efficient channel attention via softmax-normalized projection

Architecture Details

Parameter	Value
Lookback Window	96 trading days
Prediction Horizon	5 trading days
Hidden Dim (d_model)	64
FFN Dim (d_ff)	128
Attention Heads	8
Encoder Layers	2
Patch Length	16
Stride	8
Token Blend Size	2
EMA Alpha	0.9
Dynamic Projection Rank	8
Dropout	0.3

Test Results

Metric	Value
MSE	0.039362
MAE	0.132980
RMSE	0.198398

Stocks Covered

RELIANCE, TCS, HDFCBANK, INFY, ICICIBANK, HINDUNILVR, SBIN, BHARTIARTL, KOTAKBANK, ITC, LT, AXISBANK, BAJFINANCE, ASIANPAINT, MARUTI, HCLTECH, SUNPHARMA, TITAN, ULTRACEMCO, WIPRO

Training Details

• Data: NIFTY 50 stocks via yfinance (Close prices)
• Split: 70% train / 15% validation / 15% test
• Optimizer: Adam with cosine LR decay + linear warmup
• Loss: Signal Decay Loss (l^{-1} weighting)
• Learning Rate: 0.0001
• Epochs: 100 (with early stopping, patience=10)
• Batch Size: 128

Usage

import torch, json
from card_nse_predictor import CARD, CARDConfig

# Load config
with open("config.json") as f:
    cfg = CARDConfig(**json.load(f))

# Load model
model = CARD(cfg)
model.load_state_dict(torch.load("model.pt", map_location="cpu"))
model.eval()

# Input: (batch, n_stocks, lookback_days) — raw close prices
# Output: (batch, n_stocks, pred_len) — predicted close prices
x = torch.randn(1, 20, 96)
pred = model(x)
print(pred.shape)  # (1, 20, 5)

Citation

@inproceedings{xue2024card,
  title={CARD: Channel Aligned Robust Blend Transformer for Time Series Forecasting},
  author={Xue, Wang and Zhou, Tian and Wen, Qingsong and Gao, Jinyang and Ding, Bolin and Jin, Rong},
  booktitle={International Conference on Learning Representations},
  year={2024}
}