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(today, memory, horizon, ...)
  12         {
  13             # (re)initialize computed parameters
  14             private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
  15
  16             # Determine indices of no-NAs days followed by no-NAs tomorrows
  17             fdays = private$.data$getCoupleDays(max(today-memory,1), today-1)
  18
  19             # Get optional args
  20             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "mix") #or "endo", or "exo"
  21             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  22             if (hasArg(h_window))
  23             {
  24                 return ( private$.predictShapeAux(
  25                     fdays, today, horizon, list(...)$h_window, kernel, simtype, TRUE) )
  26             }
  27
  28             # Indices of similar days for cross-validation; TODO: 45 = magic number
  29             sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
  30
  31             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  32             errorOnLastNdays = function(h, kernel, simtype)
  33             {
  34                 error = 0
  35                 nb_jours = 0
  36                 for (i in intersect(fdays,sdays))
  37                 {
  38                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  39                     prediction = private$.predictShapeAux(fdays, i, horizon, h, kernel, simtype, FALSE)
  40                     if (!is.na(prediction[1]))
  41                     {
  42                         nb_jours = nb_jours + 1
  43                         error = error +
  44                             mean((private$.data$getCenteredSerie(i+1)[1:horizon] - prediction)^2)
  45                     }
  46                 }
  47                 return (error / nb_jours)
  48             }
  49
  50             if (simtype != "endo")
  51             {
  52                 h_best_exo = optimize(
  53                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="exo")$minimum
  54             }
  55             if (simtype != "exo")
  56             {
  57                 h_best_endo = optimize(
  58                     errorOnLastNdays, c(0,10), kernel=kernel, simtype="endo")$minimum
  59             }
  60
  61             if (simtype == "endo")
  62             {
  63                 return (private$.predictShapeAux(
  64                     fdays, today, horizon, h_best_endo, kernel, "endo", TRUE))
  65             }
  66             if (simtype == "exo")
  67             {
  68                 return (private$.predictShapeAux(
  69                     fdays, today, horizon, h_best_exo, kernel, "exo", TRUE))
  70             }
  71             if (simtype == "mix")
  72             {
  73                 h_best_mix = c(h_best_endo,h_best_exo)
  74                 return(private$.predictShapeAux(
  75                     fdays, today, horizon, h_best_mix, kernel, "mix", TRUE))
  76             }
  77         }
  78     ),
  79     private = list(
  80         # Precondition: "today" is full (no NAs)
  81         .predictShapeAux = function(fdays, today, horizon, h, kernel, simtype, final_call)
  82         {
  83             fdays = fdays[ fdays < today ]
  84             # TODO: 3 = magic number
  85             if (length(fdays) < 3)
  86                 return (NA)
  87
  88             data = private$.data #shorthand
  89
  90             if (simtype != "exo")
  91             {
  92                 h_endo = ifelse(simtype=="mix", h[1], h)
  93
  94                 # Distances from last observed day to days in the past
  95                 distances2 = rep(NA, length(fdays))
  96                 for (i in seq_along(fdays))
  97                 {
  98                     delta = data$getCenteredSerie(today) - data$getCenteredSerie(fdays[i])
  99                     # Require at least half of non-NA common values to compute the distance
 100                     if (sum(is.na(delta)) <= 0) #length(delta)/2)
 101                         distances2[i] = mean(delta^2) #, na.rm=TRUE)
 102                 }
 103
 104                 sd_dist = sd(distances2)
 105                 if (sd_dist < .Machine$double.eps)
 106                     sd_dist = 1 #mostly for tests... FIXME:
 107                 simils_endo =
 108                     if (kernel=="Gauss")
 109                         exp(-distances2/(sd_dist*h_endo^2))
 110                     else { #Epanechnikov
 111                         u = 1 - distances2/(sd_dist*h_endo^2)
 112                         u[abs(u)>1] = 0.
 113                         u
 114                     }
 115             }
 116
 117             if (simtype != "endo")
 118             {
 119                 h_exo = ifelse(simtype=="mix", h[2], h)
 120
 121                 M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
 122                 M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 123                 for (i in seq_along(fdays))
 124                     M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
 125
 126                 sigma = cov(M) #NOTE: robust covariance is way too slow
 127                 sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 128
 129                 # Distances from last observed day to days in the past
 130                 distances2 = rep(NA, nrow(M)-1)
 131                 for (i in 2:nrow(M))
 132                 {
 133                     delta = M[1,] - M[i,]
 134                     distances2[i-1] = delta %*% sigma_inv %*% delta
 135                 }
 136
 137                 sd_dist = sd(distances2)
 138                 simils_exo =
 139                     if (kernel=="Gauss")
 140                         exp(-distances2/(sd_dist*h_exo^2))
 141                     else { #Epanechnikov
 142                         u = 1 - distances2/(sd_dist*h_exo^2)
 143                         u[abs(u)>1] = 0.
 144                         u
 145                     }
 146             }
 147
 148             similarities =
 149                 if (simtype == "exo")
 150                     simils_exo
 151                 else if (simtype == "endo")
 152                     simils_endo
 153                 else #mix
 154                     simils_endo * simils_exo
 155
 156             prediction = rep(0, horizon)
 157             for (i in seq_along(fdays))
 158                 prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
 159             prediction = prediction / sum(similarities, na.rm=TRUE)
 160
 161             if (final_call)
 162             {
 163                 private$.params$weights <- similarities
 164                 private$.params$indices <- fdays
 165                 private$.params$window <-
 166                     if (simtype=="endo") {
 167                         h_endo
 168                     } else if (simtype=="exo") {
 169                         h_exo
 170                     } else { #mix
 171                         c(h_endo,h_exo)
 172                     }
 173             }
 174
 175             return (prediction)
 176         }
 177     )
 178 )