Semester-Level Time Series Forecasting Project: Bitcoin Prices

Overview

• End-to-end, professor-ready time-series project using a real-world OHLCV dataset (Bitcoin.csv).
• Target variable: Close. Notebook implements data preprocessing, extensive EDA, stationarity diagnostics, feature engineering, time-aware train/test split, ARIMA/SARIMA/Auto-ARIMA training, evaluation (MAE, RMSE, MAPE), and best-model persistence.

Features

• Clean, modular notebook structure suitable for academic evaluation and reuse.
• Robust preprocessing with time-aware interpolation, duplicate removal, outlier handling, optional log transform.
• High-density EDA producing many plots with concise bullet insights per figure.
• Calendar and seasonality analysis: month-of-year, day-of-week, heatmaps, decomposition.
• Stationarity diagnostics: ADF tests and visual comparisons before/after transformations.
• Feature engineering: lag features, rolling statistics, EMAs, volatility proxies.
• Time-aware train/test split (no shuffle) with clear visualization.
• Multiple forecasting baselines and models: Naive, ARIMA, SARIMA, Auto-ARIMA-style grid search.
• Model evaluation using MAE, RMSE, MAPE, with comparative table and best-model selection.
• Model persistence: save best model to best_model.pkl and reload verification.
• Practical runtime controls: frequency setting (asfreq('D')), warning suppression, tunable grid sizes.

Files

• Bitcoin.csv: dataset with Date, Open, High, Low, Close, Volume.
• TimeSeries_Forecasting_Project.ipynb: main notebook with the complete workflow.
• best_model.pkl: generated model artifact (ignored by git by default).

Quickstart (Windows)

• Create a virtual environment:
  • python -m venv .venv
  • .\.venv\Scripts\activate
• Install dependencies:
  • pip install pandas numpy matplotlib seaborn statsmodels scikit-learn notebook
• Launch Jupyter:
  • jupyter notebook
• Open and run TimeSeries_Forecasting_Project.ipynb top-to-bottom.

What the Notebook Covers

• Data Loading & Understanding: head/tail/shape/info/describe, time range, inferred frequency, target.
• Preprocessing: datetime index, missing handling (time-aware), duplicates, outlier treatment via winsorized returns, optional log transform, OHLCV resampling.
• EDA: trend, monthly/yearly behavior, day-of-week effects, bull/bear phases, rolling means/EMAs, volatility metrics, seasonal decomposition, seasonal heatmaps, spike/crash detection, cumulative returns, transformations, multivariate relations (Open vs Close, Volume vs Price), correlation and lag analysis.
• Stationarity & Diagnostics: ADF tests on original/log/differenced series, interpretation, visual checks.
• Feature Engineering: lags (1/7/14/30), rolling stats, EMAs, volatility proxy.
• Train–Test Split: 80/20 split, time-aware, visualization.
• Model Training: Naive baseline; ARIMA (AIC-based search near selected d); SARIMA (seasonal grid with inferred period); Auto-ARIMA approximation (broader AIC grid using statsmodels).
• Evaluation: MAE, RMSE, MAPE with comparison table and best model selection.
• Model Saving: persist best model to best_model.pkl, reload example and forecast verification.

Tips

• Grid searches can be computationally intensive; adjust p,q,P,Q ranges or reduce training window for speed.
• Warnings related to inferred frequency and convergence are suppressed for readability; frequency is set via asfreq('D') on the daily series.
• Optional: for pmdarima true auto_arima, install pip install pmdarima and integrate if desired.

Reproducibility

• The workflow is deterministic on fixed data; models use maximum-likelihood estimation which may vary slightly with solver tolerances. For faster consistent runs, constrain search grids.

License

• No license specified. Add one if you plan to publish.