trnnick / nnfor Goto Github PK

Hi,

When using the mlp function and setting difforder=0, the resulting model doesn't include differencing, however, the output indicates the difforder is NULL not 0. This then creates issues when fitting the same model to new data as difforder is set as NULL and not 0.

Please see the code example:

X = rnorm(500)
X = X+seq(0, 10, length.out=500)
model = nnfor::mlp(as.ts(X[1:400]),
                   xreg = cbind(1:500),
                   xreg.lags=list(0),
                   xreg.keep=list(TRUE),
                   difforder=0)

## Registered S3 method overwritten by 'quantmod':
##   method            from
##   as.zoo.data.frame zoo

#Model fits without differencing
model

## MLP fit with 5 hidden nodes and 20 repetitions.
## Univariate lags: (2)
## 1 regressor included.
## - Regressor 1 lags: (0)
## Forecast combined using the median operator.
## MSE: 1.0234.

# However, difforder is NULL
model$difforder

## NULL

model.refit = nnfor::mlp(as.ts(X),
                         model=model,
                         xreg = cbind(1:500),
                         xreg.lags = list(0),
                         xreg.keep=list(TRUE),
                         difforder=0)
#Model now includes differencing
model.refit

## MLP fit with 5 hidden nodes and 20 repetitions.
## Series modelled in differences: D1.
## Univariate lags: (2)
## 1 regressor included.
## - Regressor 1 lags: (0)
## Forecast combined using the median operator.
## MSE: 6.9452.

model.refit$difforder

## [1] 1

How do we specify type of activation function such as sigmoid, tanh in time series neural network function "mlp" in R?

I just tried to navigate to the tutorial linked in the README and received a page stating:

Error establishing a database connection

This might be out of your control, but it seemed worthwhile to report.

I want to specify both number of nodes in hidden layers and number of lags in the mlp model.
When I specify number of lags as 3 then why does it take the value of 1 as you can see in summary(fit) below.
fit <- mlp(AirPassengers,hd=c(5,2,1),lags=3)

summary(fit)
Length Class Mode
net 13 nn list
hd 3 -none- numeric
lags 1 -none- numeric
xreg.lags 0 -none- NULL
difforder 1 -none- numeric
sdummy 1 -none- logical
ff.det 1 -none- numeric
det.type 1 -none- character
y 144 ts numeric
minmax 4 -none- list
xreg.minmax 0 -none- NULL
comb 1 -none- character
fitted 140 ts numeric
MSE 1 -none- numeric
MSEH 0 -none- NULL

Also, I tried in the another dataset. When I specify both number of lags and hidden layer size then it gives following error.
Warning message:
In preprocess(y, m, lags, keep, difforder, sel.lag, allow.det.season, :
No inputs left in the network after pre-selection, forcing AR(1).
Can you please help me in this issue

I was wondering if the forecast package could be moved from "Depends" to "Imports" in the DESCRIPTION file. I am building a package around nnfor and it is giving me an error, probably because some forecast function is being overwritten by another package in my "imports" list. I found an article that states that "imports" is safer than "depends", hence this request.

https://stackoverflow.com/questions/8637993/better-explanation-of-when-to-use-imports-depends

Thanks,
Nikhil

It seems that if your ts object doesn't populated at the very beginning of your time, the error Error in apply(y, 1L, rank) : dim(X) must have a positive length will be thrown.

like this:

testdata <- ts(data = as.vector(mydata$V4), start = c(2016,3), frequency = 12)
fit <- mlp(testdata)
Error in apply(y, 1L, rank) : dim(X) must have a positive length

Error is thrown when use forecast function.
air.fit.mlp <- mlp(air.train)
air.fcst.mlp <- forecast(air.fit.mlp, h = 35)

The same problem is occurred for both local R session (MacOS) and RStudio Cloud

I get the error message from the title when running the following code:

library(tidyverse)
library(nnfor)
library(cranlogs)

rdieharder <- cran_downloads("RDieHarder", from = "2017-01-01")

train_data <- rdieharder %>%
  filter(date < as.Date("2019-01-01")) %>%
  pull(count) %>%
  ts()

# I replace this value, because this is just for testing purposes, and this value is unusually high
train_data[train_data==373] <- 3

test_data <- rdieharder %>%
  filter(date >= as.Date("2019-01-01"))

croston_model <- crost(train_data, h = 365)

nn_model <- mlp(train_data, hd.auto.type = "cv")

Same if I replace "cv" by "valid", but no issue if I add "reps = 1", or reps = "2". Any ideas?

hi,
I am doing uni variate forecasting by using SVM in r. I did my in sample forecast precisely but when i do forecast for some next time period it gives the same values. here are codes.
lagged_data<-y[complete.cases(y),]
View(lagged_data)
n<-dim(lagged_data)[1]
train_data<-lagged_data[1:(n-24),]
test_data<-lagged_data[(n-23):n,]
modelsvm = svm(inflation~.,train_data)
predYsvm = predict(modelsvm, test_data)
forecast<-data.frame(test_data$inflation, predYsvm)
till now it did good but it is in sample forecast but when i do for out sample forecast it repeats the training values.
out_predYsvm = predict(modelsvm, h=24)

It would be great to allow this to work:
mlp(BJsalesInsample,xreg=window(BJsales.lead,1,120))

Instead xreg only accepts matrix, so this needs to be done all the time:
mlp(BJsalesInsample,xreg=as.matrix(window(BJsales.lead,1,120)))

Maybe do that transformation in the code automatically if !is.matrix(xreg)?