- Only use traditional (statistical) methods if you have sufficient data. Traditional models produce accurate forecasts with at least 3 years of successive clean data.
- Consider hybrid methods of forecasting which incorporate traditional techniques as well as machine learning components to minimize chances of erratic predictions.
- Consider a three step process before concluding on your finding.
- Preprocess the data
- Statistical forecasting
- Standardize traditional forecasts using modern methods such as extreme learning machines.
- Data pre-processing will lead to more effective results. The basic items to look at while preparing your data for forecast analysis are removing outliers, interpolating missing data and normalization.
- Consider your industry. Some products and services have powerful features which vary depending on seasons. A good example is the fashion industry where color is a critical component. Some other factors that could alter your direction are:
- What is the duration of product life cycle? Products with a short term lifecycle are easy to forecast.
- Do your products vary that much? Do not combine entire sales forecasts if so?
- Is the demand stable? Consider decomposing the time series and keenly evaluate each factor on its own.
- Is it pre or post launch forecasting? Unavailability of data could render traditional techniques redundant. It might be an interesting look at user-generated content.
- Are you using pre-order or post-order data.
- Consider the predictive value of user-generated content. This can be obtained through social media and customer feedback. There is a need to look at a user as an active producer and not as a passive consumer.
- Beware of the rapidly changing environment, i.e. the rise of web 2.0, Emergence of new technology and the availability of processing power