pkg/R/F_Neighbors.R

   1 #' @include Forecaster.R
   2 #'
   3 #' @title Neighbors Forecaster
   4 #'
   5 #' @description Predict tomorrow as a weighted combination of "futures of the past" days.
   6 #'   Inherits \code{\link{Forecaster}}
   7 NeighborsForecaster = setRefClass(
   8     Class = "NeighborsForecaster",
   9     contains = "Forecaster",
  10
  11     methods = list(
  12         initialize = function(...)
  13         {
  14             callSuper(...)
  15         },
  16         predictShape = function(today, memory, horizon, ...)
  17         {
  18             # (re)initialize computed parameters
  19             params <<- list("weights"=NA, "indices"=NA, "window"=NA)
  20
  21             # Get optional args
  22             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "mix") #or "endo", or "exo"
  23             kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
  24             if (hasArg(h_window))
  25                 return (.predictShapeAux(fdays,today,horizon,list(...)$h_window,kernel,simtype,TRUE))
  26
  27             # HACK for test reports: complete some days with a few NAs, for nicer graphics
  28             nas_in_serie = is.na(data$getSerie(today))
  29             if (any(nas_in_serie))
  30             {
  31                 if (sum(nas_in_serie) >= length(nas_in_serie) / 2)
  32                     return (NA)
  33                 for (i in seq_along(nas_in_serie))
  34                 {
  35                     if (nas_in_serie[i])
  36                     {
  37                         #look left
  38                         left = i-1
  39                         while (left>=1 && nas_in_serie[left])
  40                             left = left-1
  41                         #look right
  42                         right = i+1
  43                         while (right<=length(nas_in_serie) && nas_in_serie[right])
  44                             right = right+1
  45                         #HACK: modify by-reference Data object...
  46                         data$data[[today]]$serie[i] <<-
  47                             if (left==0) data$data[[today]]$serie[right]
  48                             else if (right==0) data$data[[today]]$serie[left]
  49                             else (data$data[[today]]$serie[left] + data$data[[today]]$serie[right]) / 2.
  50                     }
  51                 }
  52             }
  53
  54             # Determine indices of no-NAs days followed by no-NAs tomorrows
  55             first_day = max(today - memory, 1)
  56             fdays = (first_day:(today-1))[ sapply(first_day:(today-1), function(i) {
  57                 !any(is.na(data$getSerie(i)) | is.na(data$getSerie(i+1)))
  58             }) ]
  59
  60             # Indices of similar days for cross-validation; TODO: 45 = magic number
  61             sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
  62
  63             # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
  64             errorOnLastNdays = function(h, kernel, simtype)
  65             {
  66                 error = 0
  67                 nb_jours = 0
  68                 for (i in intersect(fdays,sdays))
  69                 {
  70                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  71                     prediction = .predictShapeAux(fdays, i, horizon, h, kernel, simtype, FALSE)
  72                     if (!is.na(prediction[1]))
  73                     {
  74                         nb_jours = nb_jours + 1
  75                         error = error + mean((data$getCenteredSerie(i+1)[1:horizon] - prediction)^2)
  76                     }
  77                 }
  78                 return (error / nb_jours)
  79             }
  80
  81             if (simtype != "endo")
  82                 h_best_exo = optimize(errorOnLastNdays, c(0,10), kernel=kernel, simtype="exo")$minimum
  83             if (simtype != "exo")
  84                 h_best_endo = optimize(errorOnLastNdays, c(0,10), kernel=kernel, simtype="endo")$minimum
  85
  86             if (simtype == "endo")
  87                 return (.predictShapeAux(fdays, today, horizon, h_best_endo, kernel, "endo", TRUE))
  88             if (simtype == "exo")
  89                 return (.predictShapeAux(fdays, today, horizon, h_best_exo,  kernel, "exo",  TRUE))
  90             if (simtype == "mix")
  91             {
  92                 h_best_mix = c(h_best_endo,h_best_exo)
  93                 return (.predictShapeAux(fdays, today, horizon, h_best_mix,  kernel, "mix",  TRUE))
  94             }
  95         },
  96         # Precondition: "today" is full (no NAs)
  97         .predictShapeAux = function(fdays, today, horizon, h, kernel, simtype, final_call)
  98         {
  99             dat = data$data #HACK: faster this way...
 100
 101             fdays = fdays[ fdays < today ]
 102             # TODO: 3 = magic number
 103             if (length(fdays) < 3)
 104                 return (NA)
 105
 106             if (simtype != "exo")
 107             {
 108                 h_endo = ifelse(simtype=="mix", h[1], h)
 109
 110                 # Distances from last observed day to days in the past
 111                 distances2 = rep(NA, length(fdays))
 112                 for (i in seq_along(fdays))
 113                 {
 114                     delta = dat[[today]]$serie - dat[[ fdays[i] ]]$serie
 115                     # Require at least half of non-NA common values to compute the distance
 116                     if (sum(is.na(delta)) <= 0) #length(delta)/2)
 117                         distances2[i] = mean(delta^2) #, na.rm=TRUE)
 118                 }
 119
 120                 sd_dist = sd(distances2)
 121                 if (sd_dist < .Machine$double.eps)
 122                     sd_dist = 1 #mostly for tests... FIXME:
 123                 simils_endo =
 124                     if (kernel=="Gauss")
 125                         exp(-distances2/(sd_dist*h_endo^2))
 126                     else { #Epanechnikov
 127                         u = 1 - distances2/(sd_dist*h_endo^2)
 128                         u[abs(u)>1] = 0.
 129                         u
 130                     }
 131             }
 132
 133             if (simtype != "endo")
 134             {
 135                 h_exo = ifelse(simtype=="mix", h[2], h)
 136
 137                 M = matrix( nrow=1+length(fdays), ncol=1+length(dat[[today]]$exo) )
 138                 M[1,] = c( dat[[today]]$level, as.double(dat[[today]]$exo) )
 139                 for (i in seq_along(fdays))
 140                     M[i+1,] = c( dat[[ fdays[i] ]]$level, as.double(dat[[ fdays[i] ]]$exo) )
 141
 142                 sigma = cov(M) #NOTE: robust covariance is way too slow
 143                 sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
 144
 145                 # Distances from last observed day to days in the past
 146                 distances2 = rep(NA, nrow(M)-1)
 147                 for (i in 2:nrow(M))
 148                 {
 149                     delta = M[1,] - M[i,]
 150                     distances2[i-1] = delta %*% sigma_inv %*% delta
 151                 }
 152
 153                 sd_dist = sd(distances2)
 154                 simils_exo =
 155                     if (kernel=="Gauss")
 156                         exp(-distances2/(sd_dist*h_exo^2))
 157                     else { #Epanechnikov
 158                         u = 1 - distances2/(sd_dist*h_exo^2)
 159                         u[abs(u)>1] = 0.
 160                         u
 161                     }
 162             }
 163
 164             similarities =
 165                 if (simtype == "exo")
 166                     simils_exo
 167                 else if (simtype == "endo")
 168                     simils_endo
 169                 else #mix
 170                     simils_endo * simils_exo
 171
 172             prediction = rep(0, horizon)
 173             for (i in seq_along(fdays_indices))
 174                 prediction = prediction + similarities[i] * dat[[ fdays_indices[i]+1 ]]$serie[1:horizon]
 175             prediction = prediction / sum(similarities, na.rm=TRUE)
 176
 177             if (final_call)
 178             {
 179                 params$weights <<- similarities
 180                 params$indices <<- fdays_indices
 181                 params$window <<-
 182                     if (simtype=="endo") {
 183                         h_endo
 184                     } else if (simtype=="exo") {
 185                         h_exo
 186                     } else { #mix
 187                         c(h_endo,h_exo)
 188                     }
 189             }
 190
 191             return (prediction)
 192         }
 193     )
 194 )