Forecasting Process

• A process is a series of connected activities transforming inputs into outputs.

Forecasting Process

• Problem Definition
• Data Collection
• Data Analysis
• Model Selection and Fitting
• Model Validation
• Forecasting Model Deployment
• Monitoring Forecasting Model Performance

Problem Definition

• Includes understanding how the forecast will be used and the user's expectations.
• Key questions to address: desired forecast form, forecast horizon, forecast interval, and forecast accuracy level.

Data Collection

• Involves obtaining the relevant historical data for the variables to be forecast, including potential predictor variables.
• Relevant information collection, storage methods, and systems change and not all historical data is useful for current problems.
• During this phase, planning for future data collection and storage to preserve reliability and integrity is important.

Data Analysis

• An important preliminary step in selecting a forecasting model.
• Time series plots should be constructed and visually inspected for patterns like trends, seasonality, or cyclicity.
• Trend is the evolutionary movement (upward or downward) in the variable's value and can be long-term, dynamic, or of short duration.
• Seasonality is a time series component that repeats regularly (e.g., yearly), while cyclicity does not.
• Numerical summaries, such as standard deviation, percentiles, and autocorrelation, should be computed and evaluated.
• Unusual data points or potential outliers should be identified and flagged for further study.
• The purpose of data analysis is to gain a "feel" for the data and understand underlying patterns like trends and seasonality and suggests initial quantitative forecasting methods and models to explore.

Model Selection and Fitting

• Involves choosing one or more forecasting models and fitting them to the data.
• Fitting refers to estimating the unknown parameters, usually by least squares methods.

Model Validation

• Consists of evaluating the forecasting model to determine its likely performance in the intended application.
• Examination of the "fit" to historical data and the magnitude of forecast errors when the model is used on fresh data is required.
• Fitting Errors will always be smaller than forecast Errors
• Data Splitting: A method for validating a forecasting model before use, split into two segments fitting segment where the model is fit, and forecasting segment where forecasts are simulated.

Forecasting Model Deployment

• Involves using the fitted model and the resulting forecasts as use by the customer.
• Ensuring the customer understands how to use the model and generating timely forecasts routinely is vital.

Monitoring Forecasting Model Performance

• Should be ongoing after model deployment to ensure it performs satisfactorily.

Data for Forecasting

• Data are the raw materials for modeling and forecasting.
• Data vs Information: Data is preferred since it corresponds more to the description needed for the forecasting process, information is extracted or synthesized from this data.
• Output of forecasting can be information using data as an input.
• Data warehouse is a repository for modern organizational data (sales, transactions, etc.)
• Data warehouses store data in the cloud and organize, manipulate, and integrate data from multiple sources/systems.

Basic Functionality of Data Warehouse

• Data Extraction: Obtaining data from external sources.
• Data Transformation: Applying rules to prevent duplication and address issues like missing information (sometimes called data cleaning).
• Data Loading: Loading transformed data into the warehouse for modeling and analysis.

Data Quality Dimensions

• Accuracy
• Timeliness
• Completeness
• Representativeness
• Consistency

Data Cleaning

• Examines data to detect errors, missing values, outliers, or inconsistencies, and then correcting them.
• Errors can result from recording or transmission problems.
• Can be corrected by working with the original data source and it improves the forecasting process.

Checks Before Using Data to develop Models

• Check for missing data.
• Confirm if the data falls within an expected range.
• Check for potential outliers or unusual values.

Imputation

• Data Imputation: Correcting missing data or replacing outliers using an estimation process, replacing missing/erroneous values with "likely" values based on available information,enabling analysis with complete datasets.
• Mean value imputation: Replacing a missing value with the average of nonmissing observations.
• Stochastic mean value imputation: Adding a random variable to the mean value to capture noise or variability.
• Regression imputation: Computing the imputed value from a model predicting the missing value, the prediction model does not have to be linear.
• Hot deck imputation: Imputing values using values from similar complete observations.
• Cold deck imputation: Using information from a deck of cards not currently in current use.

Resouces for Forecasting

• Professional forecasting journals: Journal of Forecasting, International Journal of Forecasting, and Journal of Business Forecasting Methods and Systems
• Journals publishing new methodology, evaluation studies, case studies, and applications.
• Mainstream statistics and operations research journals that publish papers on forecasting: Journal of Business and Economic Statistics, Management Science, Naval Research Logistics, Operations Research, International Journal of Production Research, and Journal of Applied Statistics

Two Broad Types of Forecasting Techniques

• Qualitative Forecasting Techniques (Delphi Method): Subjective, require expert judgment, and used when there is limited historical data
• Example: Introducing new product via marketing tests, surveys, and sales performance analysis.
• Quantitative Forecasting Techniques: Use historical data and forecasting models to summarize data patterns and express statistical relationships.
• Regression Model: Use relationships between variables of interest and related predictor variables.
• Smoothing Models: Employ a simple function of previous observations.
• General Time Series Models: Employ statistical properties of historical data to specify a formal model and estimate parameters.

Form of the Forecast

• Point Estimates: Forecast almost always wrong, good to experience forecast errors
• Prediction Interval: A range useful in decision-making.
• Point Estimate: A single numerical value used to approximate an unknown population parameter, commonly used with statistics such as mean, proportion, or variance.
• Point Forecast: A single predicted value for a future observation that uses time series analysis.

Forecasting Problem Features

• Forecast Horizon: Number of future periods a forecast must be produced.
• Forecast Interval: Frequency with which new forecasts are prepared.
• Rolling and Moving horizon forecasting technique

Time series plots reveal patterns

• Trends
• Level Shifts
• Periods or cycle
• Unusual Observation
• Combination of Pattern