pkg/R/F_Neighbors.R

   1 #' Neighbors Forecaster
   2 #'
   3 #' Predict tomorrow as a weighted combination of "futures of the past" days.
   4 #' Inherits \code{\link{Forecaster}}
   5 #'
   6 NeighborsForecaster = R6::R6Class("NeighborsForecaster",
   7     inherit = Forecaster,
   8
   9     public = list(
  10         predictShape = function(data, today, memory, horizon, ...)
  11         {
  12             # (re)initialize computed parameters
  13             private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
  14
  15             # Do not forecast on days with NAs (TODO: softer condition...)
  16             if (any(is.na(data$getCenteredSerie(today))))
  17                 return (NA)
  18
  19             # Determine indices of no-NAs days followed by no-NAs tomorrows
  20             fdays = getNoNA2(data, max(today-memory,1), today-1)
  21
  22             # Get optional args
  23             local = ifelse(hasArg("local"), list(...)$local, TRUE) #same level + season?
  24             simtype = ifelse(hasArg("simtype"), list(...)$simtype, "none") #or "endo", or "exo"
  25             if (hasArg("window"))
  26             {
  27                 return ( private$.predictShapeAux(data,
  28                     fdays, today, horizon, local, list(...)$window, simtype, TRUE) )
  29             }
  30
  31             # Indices of similar days for cross-validation; TODO: 20 = magic number
  32             cv_days = getSimilarDaysIndices(today, data, limit=20, same_season=FALSE,
  33                 days_in=fdays)
  34
  35             # Optimize h : h |--> sum of prediction errors on last N "similar" days
  36             errorOnLastNdays = function(window, simtype)
  37             {
  38                 error = 0
  39                 nb_jours = 0
  40                 for (i in seq_along(cv_days))
  41                 {
  42                     # mix_strategy is never used here (simtype != "mix"), therefore left blank
  43                     prediction = private$.predictShapeAux(data, fdays, cv_days[i], horizon, local,
  44                         window, simtype, FALSE)
  45                     if (!is.na(prediction[1]))
  46                     {
  47                         nb_jours = nb_jours + 1
  48                         error = error +
  49                             mean((data$getSerie(cv_days[i]+1)[1:horizon] - prediction)^2)
  50                     }
  51                 }
  52                 return (error / nb_jours)
  53             }
  54
  55             # TODO: 7 == magic number
  56             if (simtype=="endo" || simtype=="mix")
  57             {
  58                 best_window_endo = optimize(
  59                     errorOnLastNdays, c(0,7), simtype="endo")$minimum
  60             }
  61             if (simtype=="exo" || simtype=="mix")
  62             {
  63                 best_window_exo = optimize(
  64                     errorOnLastNdays, c(0,7), simtype="exo")$minimum
  65             }
  66
  67             best_window =
  68                 if (simtype == "endo")
  69                     best_window_endo
  70                 else if (simtype == "exo")
  71                     best_window_exo
  72                 else if (simtype == "mix")
  73                     c(best_window_endo,best_window_exo)
  74                 else #none: value doesn't matter
  75                     1
  76
  77             return(private$.predictShapeAux(data, fdays, today, horizon, local,
  78                 best_window, simtype, TRUE))
  79         }
  80     ),
  81     private = list(
  82         # Precondition: "today" is full (no NAs)
  83         .predictShapeAux = function(data, fdays, today, horizon, local, window, simtype,
  84             final_call)
  85         {
  86             fdays_cut = fdays[ fdays < today ]
  87             if (length(fdays_cut) <= 1)
  88                 return (NA)
  89
  90             if (local)
  91             {
  92                 # Neighbors: days in "same season"; TODO: 60 == magic number...
  93                 fdays = getSimilarDaysIndices(today, data, limit=60, same_season=TRUE,
  94                     days_in=fdays_cut)
  95                 if (length(fdays) <= 1)
  96                     return (NA)
  97                 levelToday = data$getLevel(today)
  98                 distances = sapply(fdays, function(i) abs(data$getLevel(i)-levelToday))
  99                 #TODO: 2, 10, 3, 12 magic numbers here...
 100                 dist_thresh = 2
 101                 min_neighbs = min(10,length(fdays))
 102                 repeat
 103                 {
 104                     same_pollution = (distances <= dist_thresh)
 105                     nb_neighbs = sum(same_pollution)
 106                     if (nb_neighbs >= min_neighbs) #will eventually happen
 107                         break
 108                     dist_thresh = dist_thresh + 3
 109                 }
 110                 fdays = fdays[same_pollution]
 111                 max_neighbs = 12
 112                 if (nb_neighbs > max_neighbs)
 113                 {
 114                     # Keep only max_neighbs closest neighbors
 115                     fdays = fdays[
 116                         sort(distances[same_pollution],index.return=TRUE)$ix[1:max_neighbs] ]
 117                 }
 118                 if (length(fdays) == 1) #the other extreme...
 119                 {
 120                     if (final_call)
 121                     {
 122                         private$.params$weights <- 1
 123                         private$.params$indices <- fdays
 124                         private$.params$window <- 1
 125                     }
 126                     return ( data$getSerie(fdays[1])[1:horizon] )
 127                 }
 128             }
 129             else
 130                 fdays = fdays_cut #no conditioning
 131
 132             if (simtype == "endo" || simtype == "mix")
 133             {
 134                 # Compute endogen similarities using given window
 135                 window_endo = ifelse(simtype=="mix", window[1], window)
 136
 137                 # Distances from last observed day to days in the past
 138                 serieToday = data$getSerie(today)
 139                 distances2 = sapply(fdays, function(i) {
 140                     delta = serieToday - data$getSerie(i)
 141                     mean(delta^2)
 142                 })
 143
 144                 sd_dist = sd(distances2)
 145                 if (sd_dist < .25 * sqrt(.Machine$double.eps))
 146                 {
 147 #                   warning("All computed distances are very close: stdev too small")
 148                     sd_dist = 1 #mostly for tests... FIXME:
 149                 }
 150                 simils_endo = exp(-distances2/(sd_dist*window_endo^2))
 151             }
 152
 153             if (simtype == "exo" || simtype == "mix")
 154             {
 155                 # Compute exogen similarities using given window
 156                 window_exo = ifelse(simtype=="mix", window[2], window)
 157
 158                 M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
 159                 M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
 160                 for (i in seq_along(fdays))
 161                     M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
 162
 163                 sigma = cov(M) #NOTE: robust covariance is way too slow
 164                 # TODO: 10 == magic number; more robust way == det, or always ginv()
 165                 sigma_inv =
 166                     if (length(fdays) > 10)
 167                         solve(sigma)
 168                     else
 169                         MASS::ginv(sigma)
 170
 171                 # Distances from last observed day to days in the past
 172                 distances2 = sapply(seq_along(fdays), function(i) {
 173                     delta = M[1,] - M[i+1,]
 174                     delta %*% sigma_inv %*% delta
 175                 })
 176
 177                 sd_dist = sd(distances2)
 178                 if (sd_dist < .25 * sqrt(.Machine$double.eps))
 179                 {
 180 #                   warning("All computed distances are very close: stdev too small")
 181                     sd_dist = 1 #mostly for tests... FIXME:
 182                 }
 183                 simils_exo = exp(-distances2/(sd_dist*window_exo^2))
 184             }
 185
 186             similarities =
 187                 if (simtype == "exo")
 188                     simils_exo
 189                 else if (simtype == "endo")
 190                     simils_endo
 191                 else if (simtype == "mix")
 192                     simils_endo * simils_exo
 193                 else #none
 194                     rep(1, length(fdays))
 195             similarities = similarities / sum(similarities)
 196
 197             prediction = rep(0, horizon)
 198             for (i in seq_along(fdays))
 199                 prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
 200
 201             if (final_call)
 202             {
 203                 private$.params$weights <- similarities
 204                 private$.params$indices <- fdays
 205                 private$.params$window <-
 206                     if (simtype=="endo")
 207                         window_endo
 208                     else if (simtype=="exo")
 209                         window_exo
 210                     else if (simtype=="mix")
 211                         c(window_endo,window_exo)
 212                     else #none
 213                         1
 214             }
 215
 216             return (prediction)
 217         }
 218     )
 219 )