# Linear Forecasting with a small dataset

I am trying to get some forecast (5 years more) from a small dataset that is as follows:

year    euro
2010    17,785
2011    17,515
2012    16,880
2013    18,036
2014    18,400
2015    18,211
2016    18,382
2017    22,248


I've tried with ARIMA and auto.arima in R, for example:

However, nothing makes sense as it forecast a unique same value for the future. That's because the sample is too small.

I tried with a linear forecast in Excel and I got something that makes sense.

However, is this methodology applicable to my dataset?

• It is like 17.000.000 euros. Yes, I noticed that 2017 is higher. – XArtemesX Apr 23 at 8:40
• I have no replies to your questions. From what I understood, should I give up in getting a forecast from this data? – XArtemesX Apr 23 at 8:52

auto.arima() fits an ARIMA(0,0,0) model with a nonzero mean to your data. This means that it believes your data are independent and identically normally distributed. The optimal forecast for such a series is the expectation. Since your data are assumed to be identically distributed, this mean is identical for all future time periods. Thus the flat line.

This is not only an effect of little data, but also of the fact that any dynamics like trend are not obvious at all. Any trend would be solely driven by the last observation.

Note the large prediction intervals, which are due to the large noise in your data.

Here is what an ETS model fits, which is essentially exponential smoothing. It also does not fit a trend, but it adapts more quickly to what it believes to be a changing level. Note that the fitted multiplicative errors lead to widening prediction intervals.

If you believe that your last observation represents a new level on which your time series will stay for the future, you could go with this model. Alternatively, you could fit an ARIMA($$p$$,1,$$q$$) model, i.e., allow auto.arima() to fit any ARMA($$p$$,$$q$$) model to successive differences or increments. If so, you get an ARIMA(0,1,0) model, i.e., auto.arima() believes that increments are iid. Note that there is no "non-zero mean", i.e., the increments are assumed to be distributed with a zero mean. In other words: the forecast will start off your last observation, but predict no change.

Finally, if you believe that there truly is a trend in your data, you can fit an ARIMA($$p$$,2,$$q$$) model to your data, i.e., an ARMA model fitted to second differences. You get this:

This can go badly wrong, especially if you forecast out for five years and make any far-reaching decisions based on this forecast.

I personally don't think the initial ARIMA(0,0,0) forecast with a nonzero mean "makes no sense". It's exactly what I would expect from auto.arima(), which is an automatic piece of software designed to give reasonable forecasts over a wide range of situations.

In your case, an automatic forecasting algorithm may not be what you need at all, particularly if you have only a single series to forecast. The most interesting data point in your history is the last, and how you interpret it will determine what forecast you believe "makes sense". So the first order of business would be to understand what happened here. Did anything change in 2017 compared to previous years? Was a substitute product discontinued? Were marketing efforts increased? In each case, will the situation that led to this data point persist in the future, or be even stronger? The answers to questions like these will guide you to a more reasonable forecast.

In the meantime, I recommend the excellent free online book Forecasting: Principles and Practice (2nd ed.) by Athanasopoulos & Hyndman.

R code:

library(forecast)
euro <- ts(c(17785,17515,16880,18036,18400,18211,18382,22248),start=2010)
plot(forecast(auto.arima(euro),h=5))
plot(forecast(ets(euro),h=5))
plot(forecast(auto.arima(euro,d=1),h=5))
plot(forecast(auto.arima(euro,d=2),h=5))

• I tried also linear forecasting with excel. drive.google.com/open?id=1Uyv8ULj3DbBfniJPMFiYbKMYhVkb9aiz what do you think? – XArtemesX Apr 23 at 9:52
• To be quite honest, I would definitely trust anything that comes out of the forecast package more than Excel. The authors of the forecast package, among them Rob Hyndman, who is also somewhat active here, are extremely well known experts in the field. For Excel, for all I know some intern read the Wikipedia page and hacked something together. – Stephan Kolassa Apr 23 at 9:54
• Thank you very much! Now I need to understand the importance of the 2017 value. This is an item of national expenditure (that so far the statistics authority calculated only til 2017). What could help me understand better the future? Checking the total national budget for the following years (2018, 2019) in that field maybe could help? – XArtemesX Apr 23 at 10:00