library(ForeCA)

#amount of created components
ForeCA_components <- 2

#table trainData has predictors; and target as last column
#it will be split into 2 parts for training/validation 50%/50%


# last column is target
targetName <- tail(colnames(trainData),1)
predictorColnames <- colnames(trainData)[-ncol(trainData)]

#split into 2 parts for train/validation
rowSampleTrain <- 1:round(nrow(trainData)/2)
rowSampleValidate <- setdiff(1:nrow(trainData), rowSampleTrain)

#series matrix MUST have a full rank!!!!!!!111
#remove predictors with eigenvalue ~ 0
#oh god this code is so bad and slow, ikr
while(TRUE){
	covMatrix <- cov(as.matrix(trainData[rowSampleTrain,predictorColnames]))
	covMatrixRank <- qr(covMatrix)$rank
	if(covMatrixRank == ncol(covMatrix)){
		break
	}else{
		eigenValues <- abs(eigen(covMatrix)$values)
		eigenValuesLow <- order(eigenValues, decreasing=FALSE)[1:(ncol(covMatrix)-covMatrixRank)]
		predictorColnames[eigenValuesLow]
		eigenValues[eigenValuesLow]
		predictorColnames <- predictorColnames[-eigenValuesLow]
	}
}

forecaObj <- foreca(series = as.matrix(trainData[rowSampleTrain,predictorColnames]), n.comp = ForeCA_components, plot = TRUE)
summary(forecaObj)
dev.new()
plot(forecaObj)

scoreMatrix <- (as.matrix(trainData[rowSampleTrain, predictorColnames]) - forecaObj$center) %*% as.matrix(forecaObj$loadings)
#scoreMatrix <- forecaObj$scores #you can use this too, its already calculated in the model
componentNames <- colnames(scoreMatrix)
scoreMatrix <- as.data.frame(scoreMatrix)
scoreMatrix[,targetName] <- trainData[rowSampleTrain, targetName]

lmModel <- lm(paste(targetName, paste(componentNames,collapse=' + '),sep=' ~ '),data=scoreMatrix)
predictResultsForTrain <- predict(object = lmModel, newdata=scoreMatrix[,-ncol(scoreMatrix)])
predictResultsForTrain <- round(predictResultsForTrain)


scoreMatrixValidate <- (as.matrix(trainData[rowSampleValidate, predictorColnames]) - forecaObj$center) %*% as.matrix(forecaObj$loadings)
scoreMatrixValidate <- as.data.frame(scoreMatrixValidate)

predictResultsForValidation <- predict(object = lmModel, newdata=scoreMatrixValidate)
predictResultsForValidation <- round(predictResultsForValidation)

"Accuracy on train data"
mean(predictResultsForTrain == trainData[rowSampleTrain, targetName])
"Accuracy on validation data"
mean(predictResultsForValidation == trainData[rowSampleValidate, targetName])