{"id":1468,"date":"2017-12-11T09:18:42","date_gmt":"2017-12-11T09:18:42","guid":{"rendered":"http:\/\/kourentzes.com\/forecasting\/?p=1468"},"modified":"2017-12-11T09:18:42","modified_gmt":"2017-12-11T09:18:42","slug":"update-for-nnfor-reuse-and-retrain-models","status":"publish","type":"post","link":"https:\/\/kourentzes.com\/forecasting\/2017\/12\/11\/update-for-nnfor-reuse-and-retrain-models\/","title":{"rendered":"Update for nnfor: reuse and retrain models"},"content":{"rendered":"

A feature that was missing in the nnfor package was the ability to reuse and retrain mlp and elm model fits. This is now possible with the new arguments model<\/code> and retrain<\/code>.<\/p>\n

As an example, let us use the AirPassengers time series, with three different sample sizes and re-use and re-train the same model in various combinations.<\/p>\n

\r\nlibrary(nnfor)\r\n# Get some data\r\ny <- AirPassengers\r\ny1 <- window(y,end=c(1958,12))\r\ny2 <- window(y,end=c(1959,12))\r\ny3 <- window(y,end=c(1960,12))\r\n\r\n# Fit NN \r\nfit <- list()\r\nfit[[1]] <- mlp(y1)\r\nfit[[2]] <- mlp(y2,model=fit[[1]])\r\nfit[[3]] <- mlp(y2,model=fit[[1]],retrain=TRUE)\r\nfit[[4]] <- mlp(y3,model=fit[[1]])\r\nfit[[5]] <- mlp(y3,model=fit[[3]])\r\nfit[[6]] <- mlp(y3,model=fit[[1]],retrain=TRUE)\r\nfit[[7]] <- mlp(y3)\r\n\r\n# Get MSE and number of lags\r\nmse <- unlist(lapply(fit,function(x){x$MSE}))\r\nlags <- unlist(lapply(fit,function(x){length(x$lags)}))\r\nround(mse,2)\r\nlags\r\n<\/pre>\n
\n\r\n\r\n
Results<\/caption>\n
Model<\/th>\nSeries Sample<\/th>\nTraining Sample<\/th>\nRetrain<\/th>\nMSE<\/th>\nLags<\/th>\n<\/tr>\n<\/thead>\n
fit[[1]]<\/td>\nup to 1958<\/td>\nup to 1958<\/td>\nX<\/td>\n6.73<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[2]]<\/td>\nup to 1958<\/td>\nup to 1958<\/td>\n<\/td>\n61.25<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[3]]<\/td>\nup to 1959<\/td>\nup to 1959<\/td>\nX<\/td>\n6.68<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[4]]<\/td>\nup to 1960<\/td>\nup to 1958<\/td>\n<\/td>\n541.13<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[5]]<\/td>\nup to 1960<\/td>\nup to 1959<\/td>\n<\/td>\n260.22<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[6]]<\/td>\nup to 1960<\/td>\nup to 1960<\/td>\nX<\/td>\n12.65<\/td>\n9<\/td>\n<\/tr>\n\r\n
fit[[7]]<\/td>\nup to 1960<\/td>\nup to 1960<\/td>\nNew fit<\/td>\n7.95<\/td>\n10<\/td>\n<\/tr>\n<\/tbody><\/table><\/div>\n
As you can see, with different in-sample data and no retraining the in-sample MSE deteriorates. Using the “up to 1958” fit on the “up to 1959” and “up to 1960” samples returns MSE of 61.25 and 541.13 respectively. If we refit the network (keeping the specification fixed) the error reduces to 6.68 and 12.65 respectively. Building a new model on the “up to 1960” data finds a different lag structure (increasing the used lags from 9 to 10), resulting in an MSE of 7.95.<\/p>\n
The same arguments can be used with elm<\/code>.<\/p>\n
Related Posts<\/H3>