wallstrai

Forecast shares price using LSTM and news.

Wallstrai – Stock Price Forecast with LSTM

Wallstrai is a simple, interactive web app that lets users get quick forecasts of stock prices using historical data and a basic LSTM neural network model. Enter a stock ticker (e.g., AAPL), click "Get forecast", and view:

• Historical price charts (last 6 months view with high/low)
• LSTM-predicted prices for the next 15 business days
• Model evaluation on the last 30 days of data
• Training loss curves

The app pulls real-time historical data from Yahoo Finance and trains a small LSTM model on-the-fly for each query.

Project Overview

This Streamlit application provides an easy-to-use interface for visualizing stock price history and generating short-term forecasts using deep learning. Key features include:

• User input for any valid stock ticker
• 10 years of historical daily data download
• Min-Max scaling and sequence preparation (60-day lookback)
• LSTM model training with dropout and early stopping
• Rolling-window forecasting for the next 15 business days
• Interactive Plotly charts for historical data, forecast, backtesting (last 30 days), and training loss
• Basic company info display (name and website)

The goal is to demonstrate a full end-to-end ML workflow in a clean, shareable web app.

Tech Stack

• Frontend / Web Framework: Streamlit
• Data Fetching: yfinance (Yahoo Finance unofficial client)
• Data Processing & Visualization:
  • pandas
  • numpy
  • plotly.graph_objects (interactive charts)
• Machine Learning:
  • scikit-learn (MinMaxScaler)
  • tensorflow / keras (LSTM model, Dense layers, Dropout, EarlyStopping)
• Date Handling: datetime, pandas.tseries.offsets
• Deployment: Streamlit Community Cloud (or local / other hosts)

How AI is Used

The core AI component is a recurrent neural network (LSTM) trained specifically for each user-requested stock:

1. Data Preparation:

   • Fetch ~10 years of daily closing prices via yfinance
   • Normalize prices to [0, 1] using MinMaxScaler
   • Create sequences: 60-day input windows → predict the next day's close

2. Model Architecture:

   Sequential([
       LSTM(50, return_sequences=True, input_shape=(60, 1)),
       Dropout(0.2),
       LSTM(50, return_sequences=False),
       Dropout(0.2),
       Dense(1)
   ])

Improvements

News information can be taken to generate a comprenhensive forcast that also includes the fundamental analysis and merge it with the technical analysis.

Hyperparameter tuning — Use Optuna or Keras Tuner to find better layer sizes / learning rates per stock.

Ensemble or advanced models — Try Prophet, XGBoost, Transformer-based models (e.g. Temporal Fusion Transformer), or even pre-trained time-series models from Hugging Face.

Multi-step improvements — Add support for custom forecast periods, multi-ticker comparison, or portfolio-level forecasts.

Reliable data source — Migrate to Polygon.io (official APIs with keys) to avoid rate limits entirely.

Create Token for API.

Instructions

Knowing the US stock ticket, fill the filed and click enter.