Forecast, Automatic Routines vs. Experience

This morning, in our Time Series course, we’ve been playing with some data I got from google.ca/trends/. Actually, we’ve been playing on some old version, downloaded 18 months ago (discussed in a previous post, in French).

> urls = "http://freakonometrics.free.fr/report-headphones-2015.csv"
> report=read.table(
+ urls,skip=4,header=TRUE,sep=",",nrows=585)
> tail(report)
                    Semaine headphones
580 2015-02-08 - 2015-02-14         53
581 2015-02-15 - 2015-02-21         52
582 2015-02-22 - 2015-02-28         51
583 2015-03-01 - 2015-03-07         50
584 2015-03-08 - 2015-03-14         49
585 2015-03-15 - 2015-03-21         49

If we plot that weekly time series, we have

> plot(report[,2],type="l")

Working with weekly series is more complicated (at least to find a simple model, with only a few lags), so let us convert that series into a monthly one,

> source(
+   "http://freakonometrics.blog.free.fr/public/code/H2M.R")
> headphones=H2M(report,lang="FR",type="ts")
> plot(headphones)

Here, we did three things,

• we considered the logarithm of the series, because of the increasing variability we have on recent years,
• we removed (too) old observations
• we kept recent observations to test our models.

> X=log(headphones)
> T=as.numeric(time(X))
> idx=which((T>2006)&(T<2014.3))
> T=T[idx]
> X=as.numeric(X)[idx]
> plot(T,X,type="l")
> df=data.frame(X,T)
> reg=lm(X~T,data=df)
> abline(reg,col="red")
> Y=residuals(reg)

The residuals obtained after fitting that linear model will be our (supposed to be) stationnary time series. Let us consider various models.

• model 1

We obtained our first model be playing with the orders, this morning. Here, the model is slightly more complex,

> fit1=arima(Y,
+   order = c(1, 0, 10),
+   seasonal = list(order = c(1, 1, 0), 
+   period = 12))
> fit1
 
Call:
arima(x = Y, order = c(1, 0, 10), seasonal = list(order = c(1, 1, 0), period = 12))
 
Coefficients:
        ar1     ma1    ma2     ma3    ma4     
      0.930  -0.033  0.078  -0.101  0.365  
s.e.  0.064   0.102  0.132   0.119  0.152
   ma5    ma6     ma7    ma8    ma9   ma10
-0.217  0.216  -0.137  0.134  0.048  0.465 
 0.154  0.187   0.166  0.173  0.153  0.154 
  sar1
-0.578
 0.118
 
sigma^2 estimated as 0.00114:  log likelihood = 166,  aic = -305

If we look at the residuals, we have some white noise

We have here our first model.

• Model 2

For the second one, I was much more lazy

> library(forecast)
> auto.arima(ts(Y,start=2006,
+ frequency=12),max.p=12,max.q=12)
Series: ts(Y, start = 2006, frequency = 12) 
ARIMA(0,1,0)(0,1,1)[12]                    
 
Coefficients:
        sma1
      -0.950
s.e.   0.672
 
sigma^2 estimated as 0.00124:  log likelihood=155
AIC=-306   AICc=-306   BIC=-301

> fit2=arima(Y,
+            order = c(0, 1, 0),
+            seasonal = list(order = c(0, 1, 1), 
+            period = 12))
> plot(acf(residuals(fit2),lag=36),lwd=3)

Just to confirm that we actually have a white noise, here, use

> BP = function(h) Box.test(residuals(fit2),lag=h,
type='Box-Pierce')$p.value
> plot(1:24,Vectorize(BP)(1:24),type='b',
+ ylim=c(0,1),col="red")
> abline(h=.05,lty=2,col="blue")

One more time, we have a white noise. So this model here is valid.

• Model 3

> fit3=arima(Y,
+            order = c(12, 0, 0),
+            seasonal = list(order = c(0, 1, 1), 
+            period = 12))
> fit3
 
Call:
arima(x = Y, order = c(12, 0, 0), seasonal = list(order = c(0, 1, 1), period = 12))
 
Coefficients:
        ar1     ar2    ar3    ar4     ar5    ar6
      0.918  -0.044  0.014  0.253  -0.366  0.347
s.e.  0.112   0.148  0.145  0.135   0.142  0.151
         ar7     ar8    ar9    ar10   ar11
      -0.205  -0.030  0.395  -0.208  0.116
s.e.   0.152   0.139  0.134   0.162  0.169
        ar12    sma1
      -0.249  -0.767
s.e.   0.110   0.164

Again, we go a model similar, this morning, simply by playing with the orders. And one more time, we get a white noise,

• Model 4

For the last model, we can consider some exponential smoothing model

> HW = HoltWinters(log(headphones))

(again, on the logarithm of the series).

• Forecasting

For the first three models, use

> Y1=predict(fit1,n.ahead=36)
> Y2=predict(fit2,n.ahead=36)
> Y3=predict(fit3,n.ahead=36)

To visualise the three forecast, use, e.g.

> draw=function(fit=Y1){
+ straight_line=predict(reg,
+        newdata=data.frame(T=Tfutur))
+ mth=straight_line+fit$pred
+ sth2=fit$se^2
+ Xmy=exp(mth+sth2/2)
+ qinf=qlnorm(.05,mth,sdlog=sqrt(sth2))
+ qsup=qlnorm(.95,mth,sqrt(sth2))
+ 
+ plot(headphones,type="l",xlim=c(2006,2016),
+      ylim=c(25,105))
+ polygon(c(Tfutur,rev(Tfutur)),
+         c(qinf,rev(qsup)),
+         col="yellow",border=NA)
+ lines(Tfutur,Xmy,col="red")
+ return(Xmy)
+ }

Here, we have

> draw(Y1)
Time Series:
Start = 100 
End = 135 
Frequency = 1 
 [1]  45.6  46.3  47.6  52.1  52.7  52.9  68.3  
 [8]  88.8  58.5  51.2  49.5  47.4  47.8  49.4  
[15]  51.6  54.6  54.4  55.2  73.1  95.9  64.2  
[22]  56.8  54.5  52.3  53.4  54.8  56.9  61.5  
[29]  62.0  62.6  81.8 106.9  71.0  62.6  60.4  
[35]  57.9

> draw(Y2)

> draw(Y3)

And finally, for the fourth model we can use

> Y4=predict(HW,n.ahead=36,
+    prediction.interval = TRUE)
> sth2=((Y4[,"upr"]-Y4[,"fit"])/1.96)^2
> mth=Y4[,"fit"]

> Xmy=exp(mth+sth2/2)
> XHW=Xmy
> qinf=qlnorm(.05,mth,sdlog=sqrt(sth2))
> qsup=qlnorm(.95,mth,sqrt(sth2))

> plot(headphones,type="l",xlim=c(2006,2016),
+      ylim=c(25,105))
> polygon(c(Tfutur,rev(Tfutur)),
+         c(qinf,rev(qsup)),
+         col="yellow",border=NA)
> lines(Tfutur,Xmy,col="red")

• Model comparison

Here, we can compare the forecast with actual values

> X=headphones
> T=as.numeric(time(X))
> idx=which((T>=2014.3))
> T=T[idx]
> X=as.numeric(X)[idx]

> D=data.frame(T=T,X=X,X1=draw(Y1)[1:length(T)],X2=draw(Y2)[1:length(T)],X3=draw(Y3)[1:length(T)],X4=XHW[2:(1+length(T))])
> D
      T    X   X1   X2   X3   X4
1  2014 44.8 45.6 43.7 44.3 49.0
2  2014 45.2 46.3 44.3 44.9 49.6
3  2014 46.8 47.6 45.5 45.6 51.2
4  2015 49.3 52.1 47.7 48.8 53.8
5  2015 49.8 52.7 47.9 49.7 53.6
6  2015 50.9 52.9 49.0 50.8 54.4
7  2015 69.8 68.3 62.3 64.9 72.3
8  2015 91.2 88.8 82.5 85.4 92.4
9  2015 57.5 58.5 58.0 59.4 59.8
10 2015 52.5 51.2 51.9 53.6 54.9
11 2015 49.3 49.5 49.3 51.4 52.6

We can plot the errors,

> barplot(D$X1-D$X,ylim=c(-7,5))

> barplot(D$X2-D$X,ylim=c(-7,5))

x

> barplot(D$X3-D$X,ylim=c(-7,5))

> barplot(D$X4-D$X,ylim=c(-7,5))

To compare the four model, compute the sum of the squares of those errors,

> L2_dist=function(var){
+   return(sum((D[,"X"]-D[,var])^2))
+ }
 
> sapply(paste("X",1:4,sep=""),L2_dist)
   X1    X2    X3    X4 
 33.2 145.3  68.0 133.9

Model 2 and Model 4 were obtained using automatic routines. One on ARIMA models, the other one on the class of exponential smoothing models. The interesting point is thatthe we’ve been able to obtain two better models, just by playing a litle bit. From an initial guess, we look at the residuals, try a another one, and again another one, until the residuals series is a white noise… Somehow I like this idea, that I can still beat automatic routines, just using experience I have, now, on time series.