pkg/R/F_Neighbors2.R

   1 #' @include Forecaster.R
   2 #'
   3 #' Neighbors2 Forecaster
   4 #'
   5 #' Predict tomorrow as a weighted combination of "futures of the past" days.
   6 #' Inherits \code{\link{Forecaster}}
   7 #'
   8 Neighbors2Forecaster = R6::R6Class("Neighbors2Forecaster",
   9     inherit = Forecaster,
  10
  11     public = list(
  12         predictShape = function(data, today, memory, horizon, ...)
  13         {
  14             # (re)initialize computed parameters
  15             private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
  16
  17             # Do not forecast on days with NAs (TODO: softer condition...)
  18             if (any(is.na(data$getCenteredSerie(today))))
  19                 return (NA)
  20
  21             # Determine indices of no-NAs days followed by no-NAs tomorrows
  22             # Indices of similar days for cross-validation; TODO: 45 = magic number
  23             fdays = intersect(
  24                 getNoNA2(data, max(today-memory,1), today-1)
  25                 getSimilarDaysIndices(today, limit=45, same_season=TRUE) )
  26
  27             # Get optional args
  28             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  29             if (hasArg(h_window))
  30             {
  31                 return ( private$.predictShapeAux(data,
  32                     fdays, today, horizon, list(...)$h_window, kernel, TRUE) )
  33             }
  34
  35             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  36             errorOnLastNdays = function(h, kernel)
  37             {
  38                 error = 0
  39                 nb_jours = 0
  40                 for (day in fdays)
  41                 {
  42                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  43                     prediction = private$.predictShapeAux(data,fdays,day,horizon,h,kernel,FALSE)
  44                     if (!is.na(prediction[1]))
  45                     {
  46                         nb_jours = nb_jours + 1
  47                         error = error +
  48                             mean((data$getSerie(i+1)[1:horizon] - prediction)^2)
  49                     }
  50                 }
  51                 return (error / nb_jours)
  52             }
  53
  54             # h :: only for endo in this variation
  55             h_best = optimize(errorOnLastNdays, c(0,10), kernel=kernel)$minimum
  56             return (private$.predictShapeAux(data,fdays,today,horizon,h_best,kernel,TRUE))
  57         }
  58     ),
  59     private = list(
  60         # Precondition: "today" is full (no NAs)
  61         .predictShapeAux = function(data, fdays, today, horizon, h, kernel, final_call)
  62         {
  63             fdays = fdays[ fdays < today ]
  64             # TODO: 3 = magic number
  65             if (length(fdays) < 3)
  66                 return (NA)
  67
  68             # ENDO:: Distances from last observed day to days in the past
  69             distances2 = rep(NA, length(fdays))
  70             for (i in seq_along(fdays))
  71             {
  72                 delta = data$getSerie(today) - data$getSerie(fdays[i])
  73                 # Require at least half of non-NA common values to compute the distance
  74                 if ( !any( is.na(delta) ) )
  75                     distances2[i] = mean(delta^2)
  76             }
  77
  78             sd_dist = sd(distances2)
  79             if (sd_dist < .Machine$double.eps)
  80             {
  81 #                   warning("All computed distances are very close: stdev too small")
  82                 sd_dist = 1 #mostly for tests... FIXME:
  83             }
  84             simils_endo =
  85                 if (kernel=="Gauss")
  86                     exp(-distances2/(sd_dist*h_endo^2))
  87                 else
  88                 {
  89                     # Epanechnikov
  90                     u = 1 - distances2/(sd_dist*h_endo^2)
  91                     u[abs(u)>1] = 0.
  92                     u
  93                 }
  94
  95
  96             # EXOGENS: distances computations are enough
  97             # TODO: search among similar concentrations....... at this stage ?!
  98             M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
  99             M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 100             for (i in seq_along(fdays))
 101                 M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
 102
 103             sigma = cov(M) #NOTE: robust covariance is way too slow
 104             sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 105
 106             # Distances from last observed day to days in the past
 107             distances2 = rep(NA, nrow(M)-1)
 108             for (i in 2:nrow(M))
 109             {
 110                 delta = M[1,] - M[i,]
 111                 distances2[i-1] = delta %*% sigma_inv %*% delta
 112             }
 113
 114             ppv <- sort(distances2, index.return=TRUE)$ix[1:10] #..............
 115 #PPV pour endo ?
 116
 117
 118             similarities =
 119                 if (simtype == "exo")
 120                     simils_exo
 121                 else if (simtype == "endo")
 122                     simils_endo
 123                 else #mix
 124                     simils_endo * simils_exo
 125
 126             prediction = rep(0, horizon)
 127             for (i in seq_along(fdays))
 128                 prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
 129             prediction = prediction / sum(similarities, na.rm=TRUE)
 130
 131             if (final_call)
 132             {
 133                 private$.params$weights <- similarities
 134                 private$.params$indices <- fdays
 135                 private$.params$window <- h
 136             }
 137
 138             return (prediction)
 139         }
 140     )
 141 )