regression_models

1. Forex Price Direction Predictor with XGBoost:

A time series machine learning project that uses historical EUR/USD data to predict the direction of price movement using XGBoost regression and technical features like lagged close prices, smoothed averages, and noise filtering.

Project Objective

The goal of this project is to:

1. Predict the next day's closing price of EUR/USD based on past prices.

2. Evaluate model performance not just on error metrics (MSE, MAE), but more importantly:

3. Measure directional accuracy — i.e., how often the model correctly predicts if the price will go up or down.

Data Preparation:

1. Raw CSV with Date, Open, High, Low, Close is loaded.

2. Dates are parsed and formatted properly.

3. Close prices are smoothed with a 3-day moving average to reduce noise.

4. Outlier days with extreme % changes are dropped.

Feature Engineering:

Lag features are created from smoothed closing prices (e.g., lag_1, lag_2, lag_3).

These lag features serve as the model input.

Model Training:

The model used is XGBRegressor, tuned manually for best directional performance.

Data is split into train/test sets using chronological order (no shuffling).

Evaluation Metrics:

MSE / MAE: Traditional regression errors.

Directional Accuracy: The percentage of times the model correctly predicts if the next price moves up or down.

Results:

Predictions and actuals are saved to a CSV.

Optional plotting is available to visualize predictions vs actuals.

2. Coca-Cola Stock Price Prediction:

This project uses LSTM (Long Short-Term Memory) and XGBoost (Extreme Gradient Boosting) to predict daily closing prices of Coca-Cola stock using historical market data.

File descriptions:

lstm_coca_cola.py | Predicts stock prices using an LSTM neural network

xgb_coca_cola.py | Predicts stock prices using XGBoost regression

coca_cola_stock.csv | Historical Coca-Cola stock data (Date, Open, High, Low, Close, Adj Close, Volume)

LSTM Model

Key Details:

Uses last 3 days of prices to predict the next day’s closing price

Normalizes values using MinMaxScaler

Neural network architecture:

Two LSTM layers with dropout

One dense layer for output

Trained using mean_squared_error loss and adam optimizer

XGBoost Model

Key Details:

Uses lag features (previous 3 days) for prediction

Trained with XGBRegressor

Simpler and faster than LSTM; often performs surprisingly well

Output

Both models print actual vs predicted prices for a few test samples

Also calculate:

1. MSE: Mean Squared Error

2. MAE: Mean Absolute Error

3. And generate a plot showing predicted vs actual prices