sohumgautam/lstm-stock-predictor

LSTM Stock Predictor

A bidirectional LSTM with attention mechanism for stock price and trend prediction.

Model Architecture

• Type: Bidirectional LSTM with Attention
• Input Features: 25
• Hidden Size: 90
• Layers: 3
• Outputs: 6 (price/return/trend for 1-step and 3-step ahead)

Training Results

• Best Validation Loss: 0.026681920852217784
• Final Accuracy: 1.0
• Training Epoch: 850

Usage

import torch
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(repo_id="sohumgautam/lstm-stock-predictor", filename="pytorch_model.bin")
config_path = hf_hub_download(repo_id="sohumgautam/lstm-stock-predictor", filename="config.json")
norm_stats_path = hf_hub_download(repo_id="sohumgautam/lstm-stock-predictor", filename="norm_stats.csv")

# Load and use model
# (You'll need to implement the model class locally or include it in the repo)

Model Outputs

1. price_1_step: Next period price prediction
2. return_1_step: Next period return prediction
3. trend_1_step: Next period trend (up/down) prediction
4. price_3_step: 3-period ahead price prediction
5. return_3_step: 3-period ahead return prediction
6. trend_3_step: 3-period ahead trend prediction

Important Notes

• Requires input normalization using provided stats
• Input sequence length: 16 time steps
• Features must match training configuration
• Trend outputs use sigmoid activation (threshold 0.5)