Forecasting

The objective of forecasting is to produce predictions and indications (usually demand or sales) for future periods using historical data and making several assumptions. For example, we want to forecast the monthly sales of the year +1 and we take as a reference the monthly sales of the current year. An assumption could be that the sales will not change next year and our forecast will be equal to the historical data. Other simple-to-implement assumptions can be to establish a fixed increase in sales each month or a proportional increase, for example 10%. Assumptions can become much more complex when using trend or causal models.

To carry out robust forecasts, it is important to identify the elements that have an impact on the demand and use the best technique available to estimate the impact. The answers will not be found for all the elements that have an impact on the demand, but it is necessary to make the best assumptions by complementing quantitative techniques with qualitative techniques and personal business insights.

A regression can be used to define price elasticity and, knowing that the company will change its prices, this parameter can be included in the forecast. However, if we just use historical data and apply the price elasticity effect to a price shift, we are assuming that the other elements affecting the demand do not change: the demand is stable, customers do not change their behavior or taste, competitors maintain their actual prices, and so on.

FORECASTING TECHNIQUES AND METHODS
There are several methods[1] available, and choosing among them depends on five main factors:
  • -          The life cycle phase of our product or service (the emergent, growth, or mature phase);
  • -          The type of product;
  • -          The minimum accuracy required and how much we are willing to invest in time and money;
  • -          The data availability (and the cost of obtaining more data if possible);
  • -          The scope of the forecast (short, medium, or long term).

We can group the forecasting models into three categories:

  • -     Qualitative techniques: the forecast is based on the opinion of experts (the Delphi method, brainstorming, panel consensus, etc.), the opinion of customers through surveys, or market research. It is usually applied when we do not have quantitative data (a product launch) or for forecasts of several years ahead.
  • -      Time series and projections: the forecast is based on pattern recognition about past data (seasonality, cycles, and trends) and the projection of these patterns into the future. Its accuracy is greater when the product has already entered into a steady phase (the mature phase). These methods can be more stable or more sensitive to changes; however, they cannot predict turning points based on special events and can only react more or less rapidly to a change in measured data. Examples are the moving average, exponential smoothing, and time series analysis.
  • -        Causal models: the forecast is based on the relationship between the outcome (sales) and one or more predictor variables, such as product characteristics (price, quality, and availability), competitors’ product characteristics, our strategy vs competitors’ strategy (advertising, promotions, etc.), or external factors (economic or geopolitical factors). This group includes different kinds of regressions, pricing models, and customer analytics models for predicting future behavior.


In the following chapters I will start by explaining how to perform an S-curve life cycle analysis, which is considered to be a causal model but is in part also a projection model. I think that this method should not be used alone but as the base for more accurate forecasting methods. Then I will focus on the three most-used “time series and projection” methods: moving average, exponential smoothing, and time series analysis. Finally, I will include two causal methods used in revenue management (pickup models and regressions). I will not include pricing models and customer analytics models, since they have already been included in previous chapters.

FORECAST PERFORMANCE
To choose the best method to use, first of all we should consider the kind of business and the available data. However, this information is usually not sufficient and we need to compare different methods based on their performance.
The performance is calculated by the forecast error, that is to say the difference between the real data and the forecasted data. If we have enough data, it is a good practice to use part of it to build the model (training data) and part of it to test the model (test data).
Two of the commonly used error methods are:
-        MSE (mean squared error): this is the average of the squared errors of each period;
-     MAPE (mean absolute percentage error): this is the average of the absolute percentage error of each period.





[1] For more information about different models, read this interesting article:

No comments:

Post a Comment