hi,
I am working on a research which its purpose to forecast future sales demand. I have annual data of about 27 years which my data set is obviously small. I am trying to train the model which I can forecast 6 years later's sales demand.
At first, I trained my model with annual data of each year. To clarify, Year 2000's inputs data are set with year 2000's actual sale in one row. As i trained with this technique, Everything was good and i got Rsquare of about 99%. The problem is, if I want to forecast next 6 years I have to have the input data for next 6 years. for instance currency rate for next 6 years which it will decrease my model's forecast accuracy.
I came to an idea which I could train each year with input data of the previous years. For example in training, i set currency rate of year 94 with actual sales of year 2000.
With this technique I can use year 2016's input data in order to forecast year 2022's sales.
Is this technique logical?
you can use ARMA (or ARIMA etc. model) with lag of 6.
since u want to use previous years values it should be some type of auto regressive model
Hi,
in my opinion, you can use sarima and NAR-ANN model to forecast future independent values. Based on the past lags, you can forecast next lags by using these techniques. To understand more, you can read my paper:
You can use conventional techniques like ARIMA, ARMA, and others. Neural network is a good tool for your forecasting. Two types of forecasting t such as time series which is good if you have high correlation between your attributes or traditional forecasting, you can use factors to forecast your data.
@Waldemar sorry to disagree with your opinion. choosing right variables and having access to right data will improve forecast accuracy surprisingly well
1. To forecast sales demand the next 6 years in my opinion, you should do as follows:
- Consider all the factors that affect consumption of the people such as income, wages, consumer psychology, population size, .... Economic theory pointed out these factors (or variables). You collected data according to the economic variables more as possible.
- Use a method of feature selection (for example: Principal Component Anaalysis) to pick out a few variables that replaces the initial set of all variables;
- Building the forecasting model of sales demand by regressing sold values under the selected principal components (note: every economic variables having money values must be transfered in to the variables at fixed prices);
- Forecasting the selected principal components for the next 6 years using models such as: SARIMA, ARIMA or other method and then using the built model to forecast fixed prices sales demand in the next 6 years;
- Forecasting the Price Consumer index (CPI or GDP deflator) in the next 6 years (there are many methods and it was not simple).
- Using forecasted CPI (or GDP Deflator) to calculate the sales demand at current prices in the next 6 years.
2. To increase the accuracy of forecasting, you could predict sales demand using some other methods, such as ARIMA, SARIMA, ANN or other data mining technique (but I believe in this case that it will not be as good as the method I proposed).
3. Then you test encompass of forecasts (to eliminate encompassed forecasting methods) and combine forecasts for forecasting methods that are not comparable with each other to generate the new forecast result. Certainly forecasting accuracy will be improved).
Good Luck
dear Thanh,
Thanks for your time. I have already used the first technique you have proposed. problem is with this method, we are increasing parameters which could be wrong. for instance we are adding another forecast error rate for currency rate. Also all of these forecasts are effected by CPI rate because of process of fixing currency value.
I was hoping to be able to dismiss the procedure of forecasting other variables.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence