Input

Time alignment + imputation

Transformations + outliers + plots

Decomposition + stationarity evidence

Forecast models (Holt + HW + ARIMA + ML)

Evaluation + exports + config log

Financial Time Series Project

• Loads daily FDS stock prices, sorts by date, and keeps only Date + Close.

• Forces a continuous daily timeline (creates a daily calendar), then left-joins prices onto it and fills missing Close values (LOCF) so the timeline is complete.

• Builds key time-series transforms: LogPrice and LogReturn (a differenced/log-return series).

• Flags extreme return moves using a 20-day rolling z-score and visualizes price, returns, and outliers.

• Provides stationarity evidence with ACF/PACF before and after transformation (LogPrice vs LogReturn), plus ACF/PACF tables, ADF test results, and a decomposition plot (trend/seasonal/residual).

• Trains and evaluates 4 forecasting approaches on a sequential 80/20 hold-out (most recent 20%):
  Holt, Holt-Winters, ARIMA (AIC grid search), and Tree Ensemble ML.

  • Selects the ML look-back window in a data-driven way using sequential validation on candidate windows 5, 10, 20, 40, and 60, then uses the selected look-back for final testing.

• Produces a final comparison table, MAE/RMSE scores, and exports: cleaned data, predictions, model parameters, and a combined configuration log (manual + auto params).

Full Documentation

Optimizing ML Lookback Parameter

Exports + Config Log

Evaluation