KNN.acf.lin: KNN.acf.lin

In gbonte/gbcode: Code from the handbook "Statistical foundations of machine learning"

KNN.acf.lin

Description

Multioutput KNN

Arguments

X:	training input [N,n]
Y:	training output [N,m]
X.ts:	test input [N.ts,n]
k:	min number of neighbours
dist:	type of distance: euclidean, cosine
F:	forgetting factor
C:	integer parameter which sets the maximum number of neighbours (Ck)
wta:	if TRUE a winner-takes-all strategy is used; otherwise a weigthed combination is done on the basis of the l-o-o error
Acf:	autocorrelation function of the training series
Reg:	number (>1) of null terms to regularise the mean

Details

KNN.acf.lin

Multioutput KNN for multi-step-ahed prediction. It performs a locally constant model with weighted combination of local model on the basis of the autocorrelation and partial correlation properties of the training time series.

Value

vector of N.ts predictions

Author(s)

Gianluca Bontempi [e-mail]

References

Bontempi G. Ben Taieb S. Conditionally dependent strategies for multiple-step-ahead prediction in local learning, International Journal of Forecasting Volume 27, Issue 3, July–September 2011, Pages 689–699

Examples

## Multi-step ahead time series forecasting
require(gbcode)
t=seq(0,200,by=0.1)
N<-length(t)
H<-500 ## horizon prediction
TS<-sin(t)+rnorm(N,sd=0.1)
TS.tr=TS[1:(N-H)]
N.tr<-length(TS.tr)
TS.ts<-TS[(N-H+1):N]
TS.tr=array(TS.tr,c(length(TS.tr),1))
E=MakeEmbedded(TS.tr,n=3,delay=0,hor=H,1)
X<-E$inp
Y<-E$out
N<-NROW(X)
ACF.lag<-5
Y.cont<-KNN.acf.lin(X,Y,rev(TS.tr[(N.tr-H):N.tr]),Acf=acf(TS.tr,lag.max=ACF.lag,plot=FALSE)$acf,Pacf=pacf(TS.tr,lag.max=ACF.lag,plot=FALSE)$acf,TS=TS.tr)
plot(t[(N-H+1):N],TS.ts)
lines(t[(N-H+1):N],Y.cont)

gbonte/gbcode documentation built on Feb. 27, 2024, 7:38 a.m.