I have recently completed a monograph that I think you may find of great interest.

A New Paradigm of Seasonality:

Improved Time Series Forecast Accuracy, Increased Confidence in Long-Term Forecasts, and a Universe of Unexplored Seasonal Models

I’ve included the abstract below:

ABSTRACT

The inability of human beings to perceive, conceptualize, or understand the dimension of time has created the current paradigm of time series forecasting that views seasonality as a quality of data. We ask if seasonality is present in a set of time series data, and define seasonality as a repeated pattern of peaks and troughs in the data that is always of a fixed and known frequency. This has resulted in a limited number of forecast models, a limited forecast horizon making long-range forecasts unreliable, and unchallenged assumptions about the accuracy of forecast models that lead to unjustified confidence in the forecasts.

The new paradigm of time series forecasting views seasonality as a quality of time. The question isn’t whether or not a seasonal pattern exists; the question is whether a given seasonal pattern can improve the accuracy of time series forecasts for a given dataset. This expanded paradigm introduces a universe of new forecast models that take full advantage of seasonal patterns, the ability to improve the accuracy of and confidence in long-range forecasts, and an entirely objective, accuracy-based approach to selecting the optimal forecast model for any set of time series data.

This research introduces a Moving Average Annual Seasonal Relative (MAASR) forecast model and considers two hypotheses. Hypothesis 1 is that forecasts that incorporate the MAASR seasonal patterns will be significantly more accurate than non-seasonal traditional forecasts, including the Exponential Smoothing Method (ESM), the Holt linear method, and the Autoregressive Integrated Moving Average (ARIMA) model. Hypothesis 2 is that the hybrid forecasts that combine the MAASR seasonal forecast with the traditional forecasts will be significantly more accurate than their traditional counterparts. Each hypothesis is tested by considering the accuracy of quarterly forecasts, annual forecasts, and only the fourth quarter of the annual forecast.

The research proposes 10 new seasonal models, including 5 regular, calendar-based seasonal models and 5 irregular seasonal models that are not based on the calendar. Of these, 8 different examples showed significant results across 110 individual datasets. The studies include head-to-head accuracy comparisons of at least 18 years of historical forecasts for 29 different datasets, and 8 different seasonal models. The MAASR seasonal forecasts significantly and consistently improved the accuracy of the annual forecasts, increasing confidence in long-term forecasts, and frequently improved the accuracy of quarterly forecasts.

As no existing statistical software can take advantage of these new tools, I’m hoping to partner with someone who might be interested in taking advantage of the potential of this new paradigm.

I would be happy to send you a copy of the full research if you would like to move forward.

Best regards,

Kevin B. Burk