Thanks for the comment. You can adjust how many periods you forecast into the future in this technique. In this example, I forecasted one period into the future (i.e. m=1). I can however, easily adjust what value m takes in the formula. In practice you have to pick the appropriate rows in the spreadsheet. I know this is not easy to explain in words, but hopefully this makes sense.
Other Questions: 1) Can the smoothing constants be 0 and 1? Sometimes solver yields values 0 & 1 2) For Forecasting for the first 24 periods , you have always taken m =1, and the forecast is calculated by taking Level and Trend value of the previous year. Can 'm' be 2,3 etc. If we need to give more weightage to the data of the latest years, should we take 'm' always as 1 ? 3) For Forecasting for periods, t =25,26,27 etc should we use F(24+1), (F24+2), (F24+3) as Forecast formula, m = 1,2,3 ?
This way how can I forecast for next 12 months as next 12 months there are no actual demand. So how can I get trend seasonality and forecast for next 12 months?
Thank for for the response. No, that makes perfect sense. I guess I'm just a little unsatisfied with carrying on the level and trend of time t for future projects with no variation. I also feel that the cycling of the seasonality values for t-L for anything beyond the period may be a little unreliable. Perhaps I need to look into a more advanced model to which may address trending the seasonality values as well? Would ARIMA be my next step from here?
So, the forecast Column you added right beside Seasonal, Is that for those months or for 25th onwards? If not then should we just drag the forecast data from 24th onwards to forecast till 36?
I'm forecasting for my company(8 yrs historical data), Periods, t =96. Seasonality s=12. Its better if we initiate Level by taking the average demand of the first 12 months & calculate seasonality index for 12 months.Taking 3 months avg, forecast for future months is stabilized, no seasonality left. For the forecast, for ex t =97,98 etc do I have to take F(96+1),F(96+2) for calculation of level, trend keeping m =1 or can m = 2,3 also ? More priority should be for the data of the latest years.
If I need to provide a 4 months forecast of the demand, what changes should I be making to the above? Would just taking M = 4 be sufficient? Once I reach the period 24, how do I forecast for all future periods until period 28? By replacing M by 4 would the rows for the period 21-24 hold the forecasted values for the period 25-28?
ameya pradhan Someone plz correct me if I m wrong, I think this is the U stat they are talking about- docs.oracle.com/cd/E40248_01/epm.1112/cb_statistical/frameset.htm?ch07s02s03s04.html
Hello, thank you for the excellent video, which I am using for my personal study. However, I am curious as to how you can use this forecast model to predict future behaviour? For example, if I have a series with a period of 7, how can I extend the forecast model beyond t+7? This is for a forecast of daily demand (with varying seasonality specific day) where I have months of data, but I cannot determine how to extend the model for a month of future data. Any advice would be very appreciated.
here t=3 and m=1,means we are forecasting in period 3 for period 4 (3+1). in actual literature this notation is F(t,t+m), the inside braces is subscript.
Winters' method employs a level component, a trend component, and a seasonal component at each period. It uses three weights, or smoothing parameters, to update the components at each period. Initial values for the level and trend components are obtained from a linear regression on time. Initial values for the seasonal component are obtained from a dummy-variable regression using detrended data. The Winters' method smoothing equations are: · Additive model: Lt= a (Yt - St- p) + (1- a) [Lt-1 + Tt-1] Tt = g [Lt - Lt-1] + (1 - g)Tt-1 St = d (Yt - Lt) + (1 - d) St-p t = Lt-1 + Tt-1 + St-p · Multiplicative model: Lt = a (Yt / St-p) + (1-a) [Lt-1 + Tt-1] Tt = g [Lt - Lt-1] + (1 - g)Tt-1 St = d (Yt / Lt) + (1 - d) St-p t = (Lt-1 + Tt-1) St-p where · Lt is the level at time t · a is the weight for the level · Tt is the trend at time t · g is the weight for the trend · St is the seasonal component at time t · d is the weight for the seasonal component · p is the seasonal period · Yt is the data value at time t · t is the fitted value, or one-period-ahead forecast, at time t
He used only the first 3 month to prime the model because if you look at the seasonality pattern, the seasonal pattern/cycle repeats every 3 month. Usually speaking it is better to use more data to prime the model, first 2 - 3 seasonal cycles. I use the same model at work, however as far as I can tell this is one of the older Holt-Winter models with fewer variables. If you look up research online on this topic you will easy find some newer Holt-Winter models which factor in trend line variables as well random noise control and cyclical adjustments which are all necessary to minimize forecasting error in big and important data sets
