Time Series Forecasting
Last updated
Last updated
ValQ has built-in time series forecasting capabilities that allow you to forecast based on the past data collected, without writing a single line of code. There are two ways in which you can access this powerful feature:
While creating a new plan series: From the Create a Plan pop-up, select Forecast in the Type drop-down menu and click Configure.
After creating a plan series: Select any cell on the series and click Forecast from the Actions Menu to forecast for that node.
The Forecast window gives you the configuration options for time series forecasting with which you can make predictions designed to suit your needs.
If you start forecasting while creating a new series, the forecast window does not contain the Target Node option to choose from. However, after selecting the forecast type, you will be prompted to choose the same when you want to run the forecast.
M2M(+1) and M2M(+2) distribution options get enabled when you choose two or more data series for weight distribution in the Forecast pop-up.
Following are the Forecasting Methods available:
Moving Averages
Weighted Moving Average
Linear Trend
Auto Regression-Integrated Moving Averages (ARIMA)
Single Exponential Smoothing
Holt's Method (Double Exponential Smoothing)
Holt-Winter's Method(Triple Exponential Smoothing)
The moving average method forecasts the results of Target Node for the Target Period, by calculating the Moving Average of Past Data Series period values.
A Simple Moving Average is computed by dividing the sum of all the included period values by the total number of periods.
Simple Moving Average= (a1+a2+a3+..+an)/n
where,
a1,a2,a3,.....an are the period values and
n is the total number of periods (2 years = 24 periods)
The forecasted series is predicted from the past series by applying the above computation method.
The growth component can also be added as an absolute value or as a percentage when needed and the Distribution Method can be selected as Match or M2M(+1).
In this example, forecasting is done for the node 'Total Net Profit' for the year 2023 based on the Sum of 2021 and the Sum of 2022 series with equal weights as 1 and 1 respectively and 5% growth. The Forecast Method applied is the Moving Average with 'M2M(+1)' as the Distribution Method which distributes the total to its periods based on the average distribution of all Past Series Data chosen. The below figure shows the results of forecasting done based on the above configuration and the net growth is indicated in the Results section which is 4.9%.
The Weighted Moving Average method computes the forecast results of Target Node for the Target Period based on the Weighted Moving Average of Past Data Series, their Weights and Source Periods.
Unlike a simple moving average model that assigns equal weights to all the period values, a weighted moving average allows us to assign different weights to different periods. As a result, the accuracy of the model is improved by assigning more weights to the recent periods than to the older periods which could be irrelevant.
It is computed by dividing the sum of weighted period values (product of weight and the period value) by the sum of weights.
Weighted moving average WMA= (w1*a1 + w2*a2 +...+wn*an) / (w1+w2+...+wn),
where
w1,w2,w3,....wn = weights assigned to the period values
a1,a2,a3,........an = period values
The growth component can be added as an absolute value or percentage growth when needed and the Distribution Method can be selected as Match or M2M(+1).
In this example, forecasting is done for the node 'Total Net Profit' for the year 2023 based on the Sum of 2021 and the Sum of 2022 series. Both the series have been given equal weights as 1 and 1 respectively along with 15% growth. Weights are also assigned for the period from January to June as 2 and the rest of the periods as 3. The Forecast Method applied in this example is the Weighted Moving Average with 'M2M' as the Distribution Method. It distributes the total to its periods based on the weighted moving average of all the Past Series Data. The image below shows the forecasting results based on the above configuration and the net growth is indicated as 15.1% in the Results section.
The linear Trend method applies Linear Regression to predict the Target Node for the Target Period based on Past Data Series, their Weights and Source Periods.
The linear regression method models a linear relationship between the forecast variable(dependent variable) and one or more predictor variables(independent variables). A linear regression model in its simplest form follows this equation:
y=ax+b, where,
y=forecasted variable (forecasted series)
x=predictor/input variable (past series)
a= regression co-efficient (measure of the effect of the predictor variable on the forecasted variable)
b= error term/noise
In our case, ValQ forecasts the new series based on the linear relationship model to the selected past series.
The growth component can also be included as an absolute value or a percentage increase or decrease if required and the Distribution Method can be selected as Match or M2M(+1) or M2M(+2).
In this example, forecasting is done for the node 'Total Net Profit' for the year 2023 based on the Sum of 2021 and the Sum of 2022 series with weights as 1 and 2 respectively and 10% growth. The Forecast Method applied is the Linear Trend with 'Match' as the Distribution Method. The image below shows the forecasting results based on the above configuration and the Net Growth is indicated as 8.4% in the Results section.
Clicking on the table icon at the top right of the Results section displays the same results in a tabular format.
Click Apply to apply the forecasted results to the series. Since the 'Total Net Profit' node is a complex formula, ValQ opens the Goal Seek window to select the input node you want to change.
Choose any one of the key driver nodes that you want to change, from the list displayed and click Apply.
The forecasted series appears as shown below. Here 'Total Net Profit' is forecasted for the selected period with the selected past data series and weights, by changing the 'Units Sold' node. All the changed cells are highlighted in yellow.
Learn more about Goal Seek in the next section.
The following methods of forecast(ARIMA, Single/Double/Triple Exponential Smoothing) can be used if there is a minimum of 3 years or 36 periods of data series.
To apply these forecast methods, the data in the Value Field should be arranged in descending order before loading into the model, i.e. from the most recent year to the oldest year.
Customized Weights do not apply to these exponential smoothing methods as they already assign exponentially decreasing weights in their models.
Auto Regression-Integrated Moving Average method (ARIMA) is best for a short-term planning horizon. It consists of three parts.
Auto-regression: Auto Regressive (AR) regression model is built on the autocorrelation concept, where the dependent variable depends on the past values of itself. Beta(ß) constant is used for Auto Regression.
Integrated: The integrated part of ARIMA attempts to convert the non-stationarity nature of the time-series data to a stationary one by performing prediction on the difference between any two pairs of observations rather than directly on the data itself.
A stationary series is one whose statistical properties do not vary with time or not a function of time. Stationarity in Time Series also means series without a Trend or Seasonal components.
Moving Averages: It attempts to reduce the noise in the time series data by performing some aggregation operation to the past observations in terms of residual error. The primary Alpha(α) constant is used for Moving Average prediction.
Thus it creates a 'model' for a given time series based on its past values, its lags and the lagged forecast errors and that equation can be used to forecast future values.
The distribution method is M2M by default. After selecting the 3-year data series, select ARIMA method. Alpha(α) value should be provided in the box which is between 0.01 and 0.99. You can forecast without a growth factor by using none, or you can give the growth factor as an absolute value or in percentage.
In this example, forecasting is done for the node 'Overall Result' for the year 2023 based on 2021, 2020 & 2019 actuals with the weight as 1 for all and 15% Growth. The forecast method used in this example is ARIMA and the Alpha value is given as 0.15. You can check or un-check the required data series in the Results section for a better view.
Single Exponential Smoothing method is the simplest of exponential smoothing techniques. It predicts short term baseline forecast and volatility accurately. This method uses only the level component Alpha(α) constant for forecasting. It is a ratio (or a percentage) of how much importance the model will allocate to the most recent observation compared to the importance of demand history. Hence, alpha is the smoothing factor. It is used as a Short-Term and Mid-Term planning horizon.
The distribution method used in this calculation is M2M by default. After selecting the 3 years data series, select Single Exponential Smoothing method. Alpha(α) value should be provided in the box and it must be between 0.01 and 0.99. You can forecast with no growth factor by using none, or you can give the value in percentage or as an absolute value. You can also assign weights to the data series. There is an option here to optimize the model so that you get the highest possible outcome in this forecast.
Double Exponential Smoothing method estimates both Level and Trend components of the series to forecast. Hence it uses two smoothing parameters Alpha(α) and Beta(ß) for the level and trend respectively. They have to be provided in the box given in Forecast Pop-Up. It is used for short, medium and long time planning horizon.
The distribution method used in this method is by default M2M. After the selection of 3 years data series, select Double Exponential Smoothing method and enter Alpha & Beta values. Both Alpha & Beta values must be between 0.01 and 0.99. You can enter the growth in percentage or as an absolute value if required. Also, there is an option to optimize the model which gives the highest possible outcome in this forecast that can be used to compare with other forecast values.
Triple Exponential Smoothing method introduces seasonality components also along with level and trend components for a better and more robust forecast. A Holt-Winter's model is defined by its three order parameters, alpha, beta and gamma. Alpha specifies the coefficient for level smoothing. Beta specifies the coefficient for trend smoothing. Gamma specifies the coefficient for seasonal smoothing. So, the values for Alpha, Beta & Gamma should be provided in the box given in Forecast Pop-Up. It is used as a Long-Term planning horizon.
The distribution method used in this calculation is by default M2M. After the selection of 3 years' data series, select Triple Exponential Smoothing method and enter Alpha, Beta & Gamma values in the box. Alpha, Beta & Gamma values must be between 0.01 to 0.99. You can enter the growth in percentage or as an absolute value if required. There is an option to optimize the model which gives the highest possible outcome in this forecast that can be used to compare with other forecast values.
In this example, all the above exponential forecasting methods (Single, Double and Triple) have been applied for the node 'Overall Result' for the year 2023 based on 2021, 2020 and 2019 actuals. Find below the sample parameters used for these methods.
Single Exponential Smoothing: 8% Growth; Alpha = 0.99 (optimized form),
Holt's Method: 10% growth; Alpha = 0.35, Beta = 0.25,
Holt-Winter's Method: 10% growth; Alpha = 0.33, Beta = 0.33 & Gamma = 0.33.
You can simultaneously compare the results of different methods in this way and click Apply on the desired method to create the required series.
| Options | Description | |
|---|---|---|
1
Target Node
Node for which forecast is done
2
Target Periods
Period for which forecast is done
3
Source Periods
The historical data period to be considered for the forecast
4
Grow By
Period-over-period growth component can be added here if required, either as an absolute value or in percentage. Use the dropdown to change between None, Value and Percentage.
5
Past Data Series & Weights
Choose the historical data series for forecasting along with its weightage.
6
Method
Select the desired forecasting method to be applied from the list:
Linear Trend, Moving Averages, Weighted Moving Average, ARIMA, Single Exponential Smoothing, Holt's Method & Holt-Winters Method.
7
Distribution Method
Choose the distribution method as Match or M2M.
Match is to distribute the total according to the weights of periods belonging to a single data series. M2M(+1), M2M(+2),..is used to distribute the total according to the average weights of the periods from multiple data series.
