There are several types of typical forecasting business problems: product sales forecasting, demand forecasting, diffusion of innovation forecasting, return forecasting, etc. Regardless of the type of business problem, any forecasting project encompasses the following four steps: project assessment, data preparation, model selection and model parameter estimation and forecast generation.
Project assessment starts with translation of business problem into solvable predictive analytics / forecasting counterpart and evaluation of the feasibility and success of a project implementation. In particular, project assessment phase includes:
- estimating objectives and goal attainability
- determining required resources, impactful events and constrains
- checking the correspondence between objectives (business questions) and data availability and data sufficiency
- discussing forecasts automation and forecasts integration with your planning systems
- addressing issues of possible usage of census data and/or GIS
- considering forecasting accuracy/reliability vs. degree of aggregation vs. forecasting horizon
- establishing right combination of qualitative and quantitative forecasting methods
- taking up key decisions that are derived from forecasting
- developing customized report structure
- determining required forecasting software.
- forming multidisciplinary/cross-functional project team (if necessary).
- data preparation, data preprocessing, exploratory data analysis and trend analysis
- review available data sources (as a rule, dissimilar historic data, such as sales, costs, share prices, interest rates, mortality rate, and so on) in light of project objectives and business questions and creating data set for the analysis and forecasting
- producing time series from transactional time-stamped data
- converting time series data from given sampling frequency to required one
- defining variables role, exploring their distributions, missing values, outliers and possible impactful events / interventions. Imputing / interpolating missing values
- visualizing trend and other properties of variables of interest and the relationship between time series and impactful events / interventions
- identifying and probing right time series transformation
Form the adequate class of forecasting/predictive analytics models. For example, for forecasting product sales the following questions/information can be useful:
- What customers say about their intentions to continue buying the product (in particular, this info can include positives and negatives about the product and/or the company generated by social media)
- What customers are actually doing in the market
- What customers have done in the past in the market
- How stable is an environment
Depending on the answers to these questions / available information, the class of approaches is considered: qualitative only, quantitative only, or synergy of both. If it is possible to apply quantitative approaches, then the next choice should include selection of an appropriate class of models: time series modeling, causal modeling, or simulations. For example, possible candidates for the class of time series models are:
- Moving average
- Single exponential smoothing
- Trend-adjusted exponential smoothing (HoltÂ´s)
- Trend & Seasonal adjusted exponential smoothing (WinterÂ´s)
- Seasonal adjustment, etc.
Demand forecasting can be based on product life cycle, and requires different class of models. Intermittent demand methods form another approach to demand forecasting. When available data is insufficient, or requirement is to develop "if-then" scenario, then an appropriate approach can be simulation.
Evaluating modelsÂ´ accuracy and stability. Selecting the best model and assessing modelÂ´s recommendations. Incorporating judgmental overrides into forecasting process. Identifying factors that drive discrepancy between reality and forecasts.
Creating software infrastructure for the best or several combined model(s). Setting up the report structure and reporting system. Measuring forecasts accuracy and updating data/model/mechanism of forecast generation. Incorporating forecasts in business decision making processes.