Intro to Time Series Analysis

A collection of Jupyter notebooks exploring time series analysis and forecasting, from elementary univariate analysis to comparing a wide range of forecasting models.

Notebooks • Topics Covered • Implemented Models • Usage

---

Notebooks

Notebook	Dataset	Description
elementary_analysis.ipynb	Egg sales of a local shop in Sri Lanka (30 years)	Elementary analysis of a univariate time series.
univariate_vs_multivariate.ipynb	Renewable power generation & weather conditions (2017–2021, 15-min intervals)	Forecasting GHI using multiple univariate and multivariate models.

---

Topics Covered

1. Elementary Analysis (elementary_analysis.ipynb)

• Visualization — plotting the full series and sub-segments (last 10, 5, 3 years) to identify structure
• Decomposition — additive and multiplicative decomposition using seasonal_decompose; linear interpolation of COVID-era zero values before decomposing
• Stationarity testing — Augmented Dickey-Fuller (ADF) test on the full series and on sub-segments
• Detrending — two approaches: subtracting the decomposition trend component, and subtracting the least-squares line of best fit
• Deseasonalization — confirming yearly seasonality with a monthly heatmap and ANOVA F-test, then removing the seasonal component
• Autocorrelation & Partial Autocorrelation — ACF/PACF plots across 400 lags, and on detrended data to highlight lag structure
• Smoothing — residual removal via decomposition, and rolling average smoothing with 7-day and 30-day windows

2. Univariate vs. Multivariate Forecasting (univariate_vs_multivariate.ipynb)

• Pre-processing — resampling to a full 15-minute grid, forward-fill for missing values, trimming sparse boundary periods
• Visualization & seasonality checks — GHI heatmaps across hour/day/month/year/season dimensions
• ACF/PACF — daily and yearly autocorrelation analysis, including daily-aggregated data for long-range lags
• Train/test split — 80/20 chronological split
• Rolling window and sampled-section forecasting strategies — applied consistently across all models for fair comparison

---

Implemented Models

Univariate

Model	Description
FB Prophet	Additive model with configurable daily and yearly seasonality; both fixed and rolling window strategies tested.
SARIMA	Seasonal ARIMA with order (1,0,0)(1,1,0)[96]; sampled-section rolling window predictions (full rolling too slow on 15-min data).

Multivariate

Model	Description
VAR	Vector Autoregression; tested with [GHI, isSun] and [GHI, temp, humidity, isSun]; lag order selected via AIC.
LightGBM	Gradient boosting with lag features and rolling means; tested in univariate and multivariate (temp, humidity) configurations.
Ridge Regression	Linear model with lag/rolling features and StandardScaler; multivariate with temp, humidity, isSun.
Lasso Regression	Same setup as Ridge, applied for comparison.
LSTM (pretrained + finetuned)	Two-layer LSTM pretrained on the training set and finetuned per section; Huber loss with cosine LR schedule and early stopping.

---

Usage

Requirements

pip install pandas numpy matplotlib seaborn statsmodels prophet pmdarima \
            lightgbm scikit-learn torch optuna missingno kagglehub jupyter

Running the Notebooks

# Clone the repository
git clone https://github.com/ilina-d/Intro-to-time-series-analysis.git
cd Intro-to-time-series-analysis

# Launch Jupyter
jupyter notebook

Then open any .ipynb file from the Jupyter interface in your browser. Both notebooks download their datasets automatically via kagglehub.