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 #'
   8 NeighborsForecaster = R6::R6Class("NeighborsForecaster",
   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             fdays = getNoNA2(data, max(today-memory,1), today-1)
  23
  24             # Get optional args
  25             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "mix") #or "endo", or "exo"
  26             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  27             if (hasArg(h_window))
  28             {
  29                 return ( private$.predictShapeAux(data,
  30                     fdays, today, horizon, list(...)$h_window, kernel, simtype, TRUE) )
  31             }
  32
  33             # Indices of similar days for cross-validation; TODO: 45 = magic number
  34             sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
  35
  36             cv_days = intersect(fdays,sdays)
  37             # Limit to 20 most recent matching days (TODO: 20 == magic number)
  38             cv_days = sort(cv_days,decreasing=TRUE)[1:min(20,length(cv_days))]
  39
  40             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  41             errorOnLastNdays = function(h, kernel, simtype)
  42             {
  43                 error = 0
  44                 nb_jours = 0
  45                 for (i in seq_along(cv_days))
  46                 {
  47                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  48                     prediction = private$.predictShapeAux(data,
  49                         fdays, cv_days[i], horizon, h, kernel, simtype, FALSE)
  50                     if (!is.na(prediction[1]))
  51                     {
  52                         nb_jours = nb_jours + 1
  53                         error = error +
  54                             mean((data$getCenteredSerie(cv_days[i]+1)[1:horizon] - prediction)^2)
  55                     }
  56                 }
  57                 return (error / nb_jours)
  58             }
  59
  60             if (simtype != "endo")
  61             {
  62                 h_best_exo = optimize(
  63                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="exo")$minimum
  64             }
  65             if (simtype != "exo")
  66             {
  67                 h_best_endo = optimize(
  68                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="endo")$minimum
  69             }
  70
  71             if (simtype == "endo")
  72             {
  73                 return (private$.predictShapeAux(data,
  74                     fdays, today, horizon, h_best_endo, kernel, "endo", TRUE))
  75             }
  76             if (simtype == "exo")
  77             {
  78                 return (private$.predictShapeAux(data,
  79                     fdays, today, horizon, h_best_exo, kernel, "exo", TRUE))
  80             }
  81             if (simtype == "mix")
  82             {
  83                 h_best_mix = c(h_best_endo,h_best_exo)
  84                 return(private$.predictShapeAux(data,
  85                     fdays, today, horizon, h_best_mix, kernel, "mix", TRUE))
  86             }
  87         }
  88     ),
  89     private = list(
  90         # Precondition: "today" is full (no NAs)
  91         .predictShapeAux = function(data, fdays, today, horizon, h, kernel, simtype, final_call)
  92         {
  93             fdays = fdays[ fdays < today ]
  94             # TODO: 3 = magic number
  95             if (length(fdays) < 3)
  96                 return (NA)
  97
  98             if (simtype != "exo")
  99             {
 100                 h_endo = ifelse(simtype=="mix", h[1], h)
 101
 102                 # Distances from last observed day to days in the past
 103                 serieToday = data$getSerie(today)
 104                 distances2 = sapply(fdays, function(i) {
 105                     delta = serieToday - data$getSerie(i)
 106                     mean(delta^2)
 107                 })
 108
 109                 sd_dist = sd(distances2)
 110                 if (sd_dist < .Machine$double.eps)
 111                 {
 112 #                   warning("All computed distances are very close: stdev too small")
 113                     sd_dist = 1 #mostly for tests... FIXME:
 114                 }
 115                 simils_endo =
 116                     if (kernel=="Gauss")
 117                         exp(-distances2/(sd_dist*h_endo^2))
 118                     else
 119                     {
 120                         # Epanechnikov
 121                         u = 1 - distances2/(sd_dist*h_endo^2)
 122                         u[abs(u)>1] = 0.
 123                         u
 124                     }
 125             }
 126
 127             if (simtype != "endo")
 128             {
 129                 h_exo = ifelse(simtype=="mix", h[2], h)
 130
 131                 M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
 132                 M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 133                 for (i in seq_along(fdays))
 134                     M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
 135
 136                 sigma = cov(M) #NOTE: robust covariance is way too slow
 137                 sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 138
 139                 # Distances from last observed day to days in the past
 140                 distances2 = sapply(seq_along(fdays), function(i) {
 141                     delta = M[1,] - M[i+1,]
 142                     delta %*% sigma_inv %*% delta
 143                 })
 144
 145                 sd_dist = sd(distances2)
 146                 if (sd_dist < .Machine$double.eps)
 147                 {
 148 #                   warning("All computed distances are very close: stdev too small")
 149                     sd_dist = 1 #mostly for tests... FIXME:
 150                 }
 151                 simils_exo =
 152                     if (kernel=="Gauss")
 153                         exp(-distances2/(sd_dist*h_exo^2))
 154                     else
 155                     {
 156                         # Epanechnikov
 157                         u = 1 - distances2/(sd_dist*h_exo^2)
 158                         u[abs(u)>1] = 0.
 159                         u
 160                     }
 161             }
 162
 163             similarities =
 164                 if (simtype == "exo")
 165                     simils_exo
 166                 else if (simtype == "endo")
 167                     simils_endo
 168                 else #mix
 169                     simils_endo * simils_exo
 170
 171             prediction = rep(0, horizon)
 172             for (i in seq_along(fdays))
 173                 prediction = prediction + similarities[i] * data$getCenteredSerie(fdays[i]+1)[1:horizon]
 174             prediction = prediction / sum(similarities, na.rm=TRUE)
 175
 176             if (final_call)
 177             {
 178                 private$.params$weights <- similarities
 179                 private$.params$indices <- fdays
 180                 private$.params$window <-
 181                     if (simtype=="endo")
 182                         h_endo
 183                     else if (simtype=="exo")
 184                         h_exo
 185                     else #mix
 186                         c(h_endo,h_exo)
 187             }
 188
 189             return (prediction)
 190         }
 191     )
 192 )