helloadhavan/stock-LSTM

LSTM Stock Next-Day Predictor

This model was trained on the Stock-Pile dataset. It is a custom LSTM architecture with a learned ticker embedding, trained on about 5 million rows of daily stock data covering hundreds of companies across US, European and Asian markets. Given the previous 30 trading days, it predicts the next day's Open, High, Low, Close and log-Volume. It is a learning and experimentation project, not financial advice.

Files

File	Purpose
loader.py	Model, Normalizer and load_model
config.json	Ticker ids, per-ticker normalization statistics and training logs
stock_predictor/model.pth	Trained weights
MODEL_CARD.md	Architecture, data, evaluation and limitations

Usage

Download the repo by running:

huggingface-cli download helloadhavan/stock-LSTM

import torch
from .stock-LSTM.loader import load_model, Normalizer

model = load_model("stock_predictor/model.pth", "config.json")
normalizer = Normalizer("config.json")

window = {
    "Open": open_t,
    "High": high_t,
    "Low": low_t,
    "Close": close_t,
    "Volume": torch.log1p(volume_t),
}

x = normalizer(window, "AAPL")
ticker_id = normalizer.ticker_ids["AAPL"]
ids = torch.full((x.shape[0], 1), float(ticker_id))
inputs = torch.cat([x, ids], dim=-1).unsqueeze(0)

with torch.no_grad():
    prediction = model(inputs)

result = normalizer.denormalize(prediction[0], "AAPL")
result["Volume"] = torch.expm1(result["Volume"])

Each tensor in window has shape (30,). denormalize returns a dict of prices and log-Volume, so apply torch.expm1 to get raw volume.

Notes

• Volume must be log1p-transformed before normalizing.
• Only tickers listed in config.json are supported.