user@devops:~$ cat README.md
Time Series Forecasting
# Description
Complete time series forecasting project comparing two fundamentally different approaches. Facebook Prophet: additive statistical model with multiple seasonality (daily, weekly, yearly) achieving 1.87% MAPE. LSTM (PyTorch): 2 LSTM layers (hidden size 32, sequence length 72h) + MLP regressor with Huber Loss, AdamW and ReduceLROnPlateau, achieving 15.96% MAPE. Includes multiplicative decomposition, ADF stationarity test, Prophet with custom seasonality, deep learning with early stopping, and structured comparison across 6 visualizations on two datasets (Airline Passengers 144 obs and Synthetic Multi-Seasonality 17,520 obs).
# Key features
$ Multiplicative decomposition with trend, seasonality and residual
$ Prophet with daily, weekly, yearly and custom monthly seasonality
$ Bidirectional LSTM with 2 hidden layers and Huber Loss
$ Early stopping, ReduceLROnPlateau and AdamW optimizer
$ ADF stationarity test with differencing
$ 6 visualizations: EDA, decomposition, components, forecast, training, comparison
$ Structured Prophet vs LSTM comparison with MAE, RMSE and MAPE
# Gallery
# Technologies used
