Time Series Forecasting Concepts
📌 The goal of forecasting (estimation or time series forecasting) is to understand the concept conceptually and apply it to case studies, such as price estimation.
⚙️ Forecasting is a part of data mining, alongside classification, regression, anomaly detection, clustering, and association analysis.
📊 This discussion focuses on two areas: Regression (predicting a single numeric target label based on learning) and Time Series Forecasting (predicting future target variable values based on historical data).

Data-Driven Methods in Time Series Forecasting
📌 Data-driven methods require only historical data for prediction and do not necessarily need explicit class labels.
📉 Basic data-driven methods include Naïve Forecast ($F_{n+1} = Y_n$), Simple Average, Moving Average (utilizing a 'windowing' concept based on $K$ recent data points), and Weighted Moving Average.
✨ Exponential Smoothing uses a constant $\alpha$ (typically between 0.2 and 0.4) to weigh the latest data point, requiring only the last observed value for prediction.
⏳ Key time-related terms include Period ($T$) (e.g., seconds, days, months) and Horizon ($H$) (the number of future periods being forecasted).

Model-Driven Methods in Time Series Forecasting
📌 Model-driven methods treat Time ($T$) as an independent variable (predictor) and the time series value as the dependent variable.
📈 Simple model-driven methods include Linear Regression, which models prediction based on a learned linear equation ($Y = A + B \cdot X$).
🔗 More complex methods include Polynomial Regression, Linear Regression with Seasonality, and Autoregressive Integrated Moving Average (ARIMA), which combines AR and MA concepts.

Linear Regression Application Example
✍️ The process involves calculating sums for $X$, $Y$, $X^2$, $Y^2$, and $X \cdot Y$ to determine the constants $A$ (intercept) and $B$ (slope) using specific formulas derived from the dataset.
💡 Once the equation $Y = A + B \cdot X$ is established, it can be used to predict $Y$ given a future $X$ value (e.g., predicting defect count based on temperature) or to find the required $X$ to achieve a target $Y$ (e.g., finding required temperature for a target defect rate).

Windowing Technique in Practice
⚙️ Windowing involves defining a Window Size (how many past data points to use), a Step Size (how far apart to select data points), and a Horizon (how many future steps to predict).
🛠️ In tools like RapidMiner, these parameters are crucial for structuring historical data into features and corresponding labels for training time series models.
📚 Suggested exercises involve applying these methods (including splitting data 90% training/10% testing) to stock price data and analyzing COVID-19 profit/case data to compare the accuracy of various forecasting models.

Key Points & Insights
➡️ Time series forecasting is broadly categorized into Data-Driven (relying solely on historical values) and Model-Driven (treating time as a predictor).
➡️ For data-driven methods, Exponential Smoothing is often preferred over Simple Average or Naïve methods due to its adaptive weighting ($\alpha \in [0.2, 0.4]$).
➡️ Linear Regression serves as the foundational model-driven approach, where the relationship between time ($X$) and the value ($Y$) is formalized mathematically.
➡️ Practical implementation often requires testing different windowing configurations (size, step, horizon) to find the setup that yields the lowest prediction error for a specific dataset.

📸 Video summarized with SummaryTube.com on Dec 03, 2025, 03:55 UTC