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