Forecasting II
Smoothing Techniques
In the last blog, we had discussed about the error metrics like MAD, MSE which helps us to test our model accuracy based on error rates.
(Last blog : https://tulsipatro29.medium.com/an-introduction-to-forecasting-5d293c6c9f5f )
Here, let’s start discussing about the techniques available to forecast time-series data that are stationary (no trend or seasonal effects). These techniques are referred as Smoothing Techniques as they produce forecasts based on “smoothing out” the irregular fluctuations effects in the data.
Three general techniques in smoothing includes :
i. Naive forecasting models
ii. Averaging models
iii. Exponential smoothing
i. Naive Forecasting Models
These are simple models in which it is assumed that the more recent time periods of data represents the best predictions or forecasts for future outcomes. Basically, they don’t take into account data trend, cyclical effects, or seasonality.
The simplest of the naive forecasting methods is the method in which the forecast for a given period is the value for the previous time period.
As an example, if 500 apples were sold by the retailer last week, then the naive forecasting model would predict that the retailer will sell 500 apples this week even.
ii. Averaging Models
Naive models are based on the value of one time period and the forecasts from this model becomes a function of irregular fluctuations of the data; as a result, the forecasts are “over steered”.
With Averaging models, a forecaster enters information from several time periods into the forecast and “smoothes” the data. These models are computed by averaging data from several time periods and using the average as the forecast for the next time period.
Simple Averages
Simple Average Model is the most elementary of the averaging models. The model says, the forecast for time period t is the average of the values for a given number of previous time periods,
Moving Averages
A Moving average is an average that is updated or recomputed for every new time period being considered. The most recent information is utilized in each new moving average.
Some of the disadvantages are :
i. It is difficult to choose the optimal length of time for which to compute the moving average.
ii. Moving averages do not usually adjust for time-series effects like trend, cycles or seasonality.
To determine the more optimal lengths for which to compute the moving averages, we would need to forecast with several different lengths and compare the errors produced by them.
Weighted Moving Averages
It can happen sometimes that the forecaster may want to place more weight on certain periods of time than on others. For example, he/she might believe that the previous month’s value is five times as important as the other months. A moving average where some time periods are weighted differently than others, is described as Weighted moving average.
For example, a 3-month weighted average is computed by weighting last month’s value by 3, the value for previous month by 2, and the value for month before that by 1.
This weighted average is computed as
where
M(t-1) = last month’s value
M(t-2) = value for previous month
M(t-3) = value for month before previous month
The divisor is 6. The divisor always equals the total number of weights.
The forecaster can experiment with different weights in using the weighted moving average as a technique.
The third technique “Exponential Smoothing” will be discussed further in my upcoming blog.
