Correlated daily time series and forecasting in the M4 competition

Abstract

We participated in the M4 competition for time series forecasting and describe here our methods for forecasting daily time series. We used an ensemble of five statistical forecasting methods and a method that we refer to as the correlator. Our retrospective analysis using the ground truth values published by the M4 organisers after the competition demonstrates that the correlator was responsible for most of our gains over the naive constant forecasting method. We identify data leakage as one reason for its success, partly due to test data selected from different time intervals, and partly due to quality issues in the original time series. We suggest that future forecasting competitions should provide actual dates for the time series so that some of those leakages could be avoided by the participants.

Complex exponential smoothing

“…Several time series in the competition data exhibit extreme properties (see, Darin & Stellwagen, 2020; Fildes, 2020; Ingel et al, 2020), which caused the benchmark ETS to break down. This was due to influential outliers and structural breaks in the data.…”

Grey Forecast Model with Aging Fractional Accumulation and Its Properties

“…To prove the effectiveness of the proposed aging accumulation operator, seven existing grey accumulation operators such as the traditional first-order accumulation generation operation, the damping accumulation generation operation [25], the adjacent accumulation generation operation [35], the first-order new information priority accumulation generation operation [23], the conformable fractional accumulation generation operation [36], the fractional order accumulation generation operation [22], and fractional Hausdorff accumulation generation operation [37] are compared and analyzed. e data of forecasting competition are often used to verify the performance of forecasting methods [38,39]. Take the first nine data of the N7 series in the M3 prediction contest as an example.…”

The M4 competition: Bigger. Stronger. Better

“…The first place in the hourly frequency and third in the predictions intervals was achieved by Doornik et al (2019), who developed a forecasting method that calibrates the average of a simple yet adaptive autoregressive model (Rho) and a damped trend estimation of the growth rate (Delta). The Tartu team's second place on the daily frequency (Ingel et al, 2019) was based on a combination of five statistical models as well as a simple observation that some of the daily data are highly correlated. Finally, Darin and Stellwagen (2019) achieved the first place in the weekly subset by using their proprietary software, Forecast Pro, to obtain baseline forecasts and appropriately determining a performance indicator that matched the error measure used in the competition as well as appropriately reviewing and revising cases where the baseline forecasts could be inadequate.…”