CODEWITHSUNDEEP
rneural-networktime-seriesforecastinghybrid
Chen
Chen

Reputation: 21

hybridModel of Auto.arima and ANN produce point forecast outside of 95% CI

I have been working on time series forecasting and recently read about how the hybrid model of auto.arima and ann provide better/more accurate forecasting results. I have six time series data sets, the hybrid model work wonders for five of them but it gives weird results for the other.

I ran the model using the following to packages: library(forecast) library(forecastHybrid)

Here is the data:

ts.data

     Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2012           1  16  41  65  87 104 152 203 213 263
2013 299 325 388 412 409 442 447 421 435 448 447 443
2014 454 446 467 492 525

Model:

fit <- hybridModel(ts.data, model="an")

Forecast results for the next 5 periods:

forecast(fit, 5)

   Point       Forecast    Lo 80    Hi 80    Lo 95    Hi 95
Jun 2014       594.6594 519.2914 571.0163 505.6007 584.7070
Jul 2014       702.1626 528.7327 601.8827 509.3710 621.2444
Aug 2014       738.5732 540.6665 630.2566 516.9534 653.9697
Sep 2014       752.1329 553.8905 657.3403 526.5090 684.7218
Oct 2014       762.7481 567.9391 683.5994 537.3256 714.2129

You see how the point forecasts are outside of the 95% confidence interval. Does anybody know what this is happening and how I could fix it?

Any thoughts and insights are appreciated! Thanks in advance.

Upvotes: 0

Views: 850

Answers (1)

dashaub
dashaub

Reputation: 71

See the description of this issue here
tl;dr nnetar models do not create prediction intervals, so these are not included in the ensemble prediction intervals. When the "forecast" package adds this behavior (on the road map for 2016), the prediction intervals and point forecasts will be consistent

Upvotes: 0

Related Questions