pkg/R/F_Neighbors.R

   1 #' @include Forecaster.R
   2 #'
   3 #' Neighbors Forecaster
   4 #'
   5 #' Predict tomorrow as a weighted combination of "futures of the past" days.
   6 #' Inherits \code{\link{Forecaster}}
   7 NeighborsForecaster = R6::R6Class("NeighborsForecaster",
   8     inherit = Forecaster,
   9
  10     public = list(
  11         predictShape = function(data, today, memory, horizon, ...)
  12         {
  13             # (re)initialize computed parameters
  14             private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
  15
  16             # Do not forecast on days with NAs (TODO: softer condition...)
  17             if (any(is.na(data$getCenteredSerie(today))))
  18                 return (NA)
  19
  20             # Determine indices of no-NAs days followed by no-NAs tomorrows
  21             fdays = getNoNA2(data, max(today-memory,1), today-1)
  22
  23             # Get optional args
  24             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "mix") #or "endo", or "exo"
  25             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  26             if (hasArg(h_window))
  27             {
  28                 return ( private$.predictShapeAux(data,
  29                     fdays, today, horizon, list(...)$h_window, kernel, simtype, TRUE) )
  30             }
  31
  32             # Indices of similar days for cross-validation; TODO: 45 = magic number
  33             sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
  34
  35             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  36             errorOnLastNdays = function(h, kernel, simtype)
  37             {
  38                 error = 0
  39                 nb_jours = 0
  40                 for (i in intersect(fdays,sdays))
  41                 {
  42                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  43                     prediction = private$.predictShapeAux(data,
  44                         fdays, i, horizon, h, kernel, simtype, FALSE)
  45                     if (!is.na(prediction[1]))
  46                     {
  47                         nb_jours = nb_jours + 1
  48                         error = error +
  49                             mean((data$getCenteredSerie(i+1)[1:horizon] - prediction)^2)
  50                     }
  51                 }
  52                 return (error / nb_jours)
  53             }
  54
  55             if (simtype != "endo")
  56             {
  57                 h_best_exo = optimize(
  58                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="exo")$minimum
  59             }
  60             if (simtype != "exo")
  61             {
  62                 h_best_endo = optimize(
  63                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="endo")$minimum
  64             }
  65
  66             if (simtype == "endo")
  67             {
  68                 return (private$.predictShapeAux(data,
  69                     fdays, today, horizon, h_best_endo, kernel, "endo", TRUE))
  70             }
  71             if (simtype == "exo")
  72             {
  73                 return (private$.predictShapeAux(data,
  74                     fdays, today, horizon, h_best_exo, kernel, "exo", TRUE))
  75             }
  76             if (simtype == "mix")
  77             {
  78                 h_best_mix = c(h_best_endo,h_best_exo)
  79                 return(private$.predictShapeAux(data,
  80                     fdays, today, horizon, h_best_mix, kernel, "mix", TRUE))
  81             }
  82         }
  83     ),
  84     private = list(
  85         # Precondition: "today" is full (no NAs)
  86         .predictShapeAux = function(data, fdays, today, horizon, h, kernel, simtype, final_call)
  87         {
  88             fdays = fdays[ fdays < today ]
  89             # TODO: 3 = magic number
  90             if (length(fdays) < 3)
  91                 return (NA)
  92
  93             if (simtype != "exo")
  94             {
  95                 h_endo = ifelse(simtype=="mix", h[1], h)
  96
  97                 # Distances from last observed day to days in the past
  98                 distances2 = rep(NA, length(fdays))
  99                 for (i in seq_along(fdays))
 100                 {
 101                     delta = data$getCenteredSerie(today) - data$getCenteredSerie(fdays[i])
 102                     # Require at least half of non-NA common values to compute the distance
 103                     if (sum(is.na(delta)) <= 0) #length(delta)/2)
 104                         distances2[i] = mean(delta^2) #, na.rm=TRUE)
 105                 }
 106
 107                 sd_dist = sd(distances2)
 108                 if (sd_dist < .Machine$double.eps)
 109                     sd_dist = 1 #mostly for tests... FIXME:
 110                 simils_endo =
 111                     if (kernel=="Gauss")
 112                         exp(-distances2/(sd_dist*h_endo^2))
 113                     else { #Epanechnikov
 114                         u = 1 - distances2/(sd_dist*h_endo^2)
 115                         u[abs(u)>1] = 0.
 116                         u
 117                     }
 118             }
 119
 120             if (simtype != "endo")
 121             {
 122                 h_exo = ifelse(simtype=="mix", h[2], h)
 123
 124                 M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
 125                 M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 126                 for (i in seq_along(fdays))
 127                     M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
 128
 129                 sigma = cov(M) #NOTE: robust covariance is way too slow
 130                 sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 131
 132                 # Distances from last observed day to days in the past
 133                 distances2 = rep(NA, nrow(M)-1)
 134                 for (i in 2:nrow(M))
 135                 {
 136                     delta = M[1,] - M[i,]
 137                     distances2[i-1] = delta %*% sigma_inv %*% delta
 138                 }
 139
 140                 sd_dist = sd(distances2)
 141                 simils_exo =
 142                     if (kernel=="Gauss")
 143                         exp(-distances2/(sd_dist*h_exo^2))
 144                     else { #Epanechnikov
 145                         u = 1 - distances2/(sd_dist*h_exo^2)
 146                         u[abs(u)>1] = 0.
 147                         u
 148                     }
 149             }
 150
 151             similarities =
 152                 if (simtype == "exo")
 153                     simils_exo
 154                 else if (simtype == "endo")
 155                     simils_endo
 156                 else #mix
 157                     simils_endo * simils_exo
 158
 159             prediction = rep(0, horizon)
 160             for (i in seq_along(fdays))
 161                 prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
 162             prediction = prediction / sum(similarities, na.rm=TRUE)
 163
 164             if (final_call)
 165             {
 166                 private$.params$weights <- similarities
 167                 private$.params$indices <- fdays
 168                 private$.params$window <-
 169                     if (simtype=="endo") {
 170                         h_endo
 171                     } else if (simtype=="exo") {
 172                         h_exo
 173                     } else { #mix
 174                         c(h_endo,h_exo)
 175                     }
 176             }
 177
 178             return (prediction)
 179         }
 180     )
 181 )