Time Series Data Analysis - Forecasting

 

Time series data analysis in data science involves examining data points collected or recorded at regular or sequential time intervals to uncover underlying patterns, trends, and seasonal behaviors.

Core Components of Time Series Data

• Trend: The long-term directional movement of the data (increasing, decreasing, or flat over time).

• Seasonality: Periodic and predictable fluctuations that repeat over a fixed interval (e.g., daily, monthly, or yearly).

• Cyclicity: Oscillations or wave-like patterns that occur over non-fixed, longer-term periods (such as economic cycles).

• Residuals / Noise: Unpredictable, random variations left over after removing trend and seasonal components.

Key Concepts & Workflow

1. Stationarity: Many classical models require stationarity (a constant mean, variance, and autocorrelation over time) verified via tests like the Augmented Dickey-Fuller (ADF) test. Non-stationary data is typically made stationary using differencing, logarithmic transformations, or moving averages.

2. Decomposition: Breaking a series down into trend, seasonal, and residual layers (using methods like STL decomposition).

3. Feature Engineering: Creating lag features, rolling statistics, calendar attributes, and cyclic transformations (sine/cosine encodings) for machine learning models.

4. Validation: Because time series data possesses temporal dependencies, data scientists use chronological (forward-chaining or TimeSeriesSplit) splits rather than random train-test splits to prevent data leakage.

Popular Time Series Models & Techniques

• Statistical Models:

• ARIMA / ARMA / SARIMA: Uses autoregression, differencing, and moving averages to model linear dependencies.

• Exponential Smoothing (Holt-Winters): Weights past observations exponentially, ideal for capturing level, trend, and seasonality.

• Machine Learning & Deep Learning Models:

• Tree-based Ensembles: Algorithms like XGBoost trained on engineered lag and rolling features.

• Recurrent Neural Networks (LSTM): Designed to capture long-term sequential and temporal dependencies.

https://www.ibm.com/think/topics/time-series-model

https://medium.com/@datasciencewizards/introduction-to-time-series-analysis-i-82d614d462b0

Time Series Analysis in Data Science: Forecasting Trends and Anomalies

Publisher: IEEE

Latha Narayanan Valli; N. Sujatha

Published in: 2024 5th International Conference on Smart Electronics and Communication (ICOSEC)

https://ieeexplore.ieee.org/document/10722052