| 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 | local = ifelse(hasArg("local"), list(...)$local, TRUE) #same level + season? |
| 26 | simtype = ifelse(hasArg("simtype"), list(...)$simtype, "none") #or "endo", or "exo" |
| 27 | if (hasArg("window")) |
| 28 | { |
| 29 | return ( private$.predictShapeAux(data, |
| 30 | fdays, today, horizon, local, list(...)$window, simtype, TRUE) ) |
| 31 | } |
| 32 | |
| 33 | # Indices of similar days for cross-validation; TODO: 20 = magic number |
| 34 | cv_days = getSimilarDaysIndices(today, data, limit=20, same_season=FALSE, |
| 35 | days_in=fdays) |
| 36 | |
| 37 | # Optimize h : h |--> sum of prediction errors on last N "similar" days |
| 38 | errorOnLastNdays = function(window, simtype) |
| 39 | { |
| 40 | error = 0 |
| 41 | nb_jours = 0 |
| 42 | for (i in seq_along(cv_days)) |
| 43 | { |
| 44 | # mix_strategy is never used here (simtype != "mix"), therefore left blank |
| 45 | prediction = private$.predictShapeAux(data, fdays, cv_days[i], horizon, local, |
| 46 | window, simtype, FALSE) |
| 47 | if (!is.na(prediction[1])) |
| 48 | { |
| 49 | nb_jours = nb_jours + 1 |
| 50 | error = error + |
| 51 | mean((data$getSerie(cv_days[i]+1)[1:horizon] - prediction)^2) |
| 52 | } |
| 53 | } |
| 54 | return (error / nb_jours) |
| 55 | } |
| 56 | |
| 57 | # TODO: 7 == magic number |
| 58 | if (simtype != "endo") |
| 59 | { |
| 60 | best_window_exo = optimize( |
| 61 | errorOnLastNdays, c(0,7), simtype="exo")$minimum |
| 62 | } |
| 63 | if (simtype != "exo") |
| 64 | { |
| 65 | best_window_endo = optimize( |
| 66 | errorOnLastNdays, c(0,7), simtype="endo")$minimum |
| 67 | } |
| 68 | |
| 69 | if (simtype == "endo") |
| 70 | { |
| 71 | return (private$.predictShapeAux(data, fdays, today, horizon, local, |
| 72 | best_window_endo, "endo", TRUE)) |
| 73 | } |
| 74 | if (simtype == "exo") |
| 75 | { |
| 76 | return (private$.predictShapeAux(data, fdays, today, horizon, local, |
| 77 | best_window_exo, "exo", TRUE)) |
| 78 | } |
| 79 | if (simtype == "mix") |
| 80 | { |
| 81 | return(private$.predictShapeAux(data, fdays, today, horizon, local, |
| 82 | c(best_window_endo,best_window_exo), "mix", TRUE)) |
| 83 | } |
| 84 | } |
| 85 | ), |
| 86 | private = list( |
| 87 | # Precondition: "today" is full (no NAs) |
| 88 | .predictShapeAux = function(data, fdays, today, horizon, local, window, simtype, |
| 89 | final_call) |
| 90 | { |
| 91 | fdays_cut = fdays[ fdays < today ] |
| 92 | if (length(fdays_cut) <= 1) |
| 93 | return (NA) |
| 94 | |
| 95 | if (local) |
| 96 | { |
| 97 | # Neighbors: days in "same season"; TODO: 60 == magic number... |
| 98 | fdays = getSimilarDaysIndices(today, data, limit=60, same_season=TRUE, |
| 99 | days_in=fdays_cut) |
| 100 | if (length(fdays) <= 1) |
| 101 | return (NA) |
| 102 | levelToday = data$getLevel(today) |
| 103 | distances = sapply(fdays, function(i) abs(data$getLevel(i)-levelToday)) |
| 104 | #TODO: 2, 10, 3, 12 magic numbers here... |
| 105 | dist_thresh = 2 |
| 106 | min_neighbs = min(10,length(fdays)) |
| 107 | repeat |
| 108 | { |
| 109 | same_pollution = (distances <= dist_thresh) |
| 110 | nb_neighbs = sum(same_pollution) |
| 111 | if (nb_neighbs >= min_neighbs) #will eventually happen |
| 112 | break |
| 113 | dist_thresh = dist_thresh + 3 |
| 114 | } |
| 115 | fdays = fdays[same_pollution] |
| 116 | max_neighbs = 12 |
| 117 | if (nb_neighbs > max_neighbs) |
| 118 | { |
| 119 | # Keep only max_neighbs closest neighbors |
| 120 | fdays = fdays[ |
| 121 | sort(distances[same_pollution],index.return=TRUE)$ix[1:max_neighbs] ] |
| 122 | } |
| 123 | if (length(fdays) == 1) #the other extreme... |
| 124 | { |
| 125 | if (final_call) |
| 126 | { |
| 127 | private$.params$weights <- 1 |
| 128 | private$.params$indices <- fdays |
| 129 | private$.params$window <- 1 |
| 130 | } |
| 131 | return ( data$getSerie(fdays[1])[1:horizon] ) #what else?! |
| 132 | } |
| 133 | } |
| 134 | else |
| 135 | fdays = fdays_cut #no conditioning |
| 136 | |
| 137 | if (simtype == "endo" || simtype == "mix") |
| 138 | { |
| 139 | # Compute endogen similarities using given window |
| 140 | window_endo = ifelse(simtype=="mix", window[1], window) |
| 141 | |
| 142 | # Distances from last observed day to days in the past |
| 143 | serieToday = data$getSerie(today) |
| 144 | distances2 = sapply(fdays, function(i) { |
| 145 | delta = serieToday - data$getSerie(i) |
| 146 | mean(delta^2) |
| 147 | }) |
| 148 | |
| 149 | sd_dist = sd(distances2) |
| 150 | if (sd_dist < .25 * sqrt(.Machine$double.eps)) |
| 151 | { |
| 152 | # warning("All computed distances are very close: stdev too small") |
| 153 | sd_dist = 1 #mostly for tests... FIXME: |
| 154 | } |
| 155 | simils_endo = exp(-distances2/(sd_dist*window_endo^2)) |
| 156 | } |
| 157 | |
| 158 | if (simtype == "exo" || simtype == "mix") |
| 159 | { |
| 160 | # Compute exogen similarities using given window |
| 161 | window_exo = ifelse(simtype=="mix", window[2], window) |
| 162 | |
| 163 | M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) ) |
| 164 | M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) ) |
| 165 | for (i in seq_along(fdays)) |
| 166 | M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) ) |
| 167 | |
| 168 | sigma = cov(M) #NOTE: robust covariance is way too slow |
| 169 | # TODO: 10 == magic number; more robust way == det, or always ginv() |
| 170 | sigma_inv = |
| 171 | if (length(fdays) > 10) |
| 172 | solve(sigma) |
| 173 | else |
| 174 | MASS::ginv(sigma) |
| 175 | |
| 176 | # Distances from last observed day to days in the past |
| 177 | distances2 = sapply(seq_along(fdays), function(i) { |
| 178 | delta = M[1,] - M[i+1,] |
| 179 | delta %*% sigma_inv %*% delta |
| 180 | }) |
| 181 | |
| 182 | sd_dist = sd(distances2) |
| 183 | if (sd_dist < .25 * sqrt(.Machine$double.eps)) |
| 184 | { |
| 185 | # warning("All computed distances are very close: stdev too small") |
| 186 | sd_dist = 1 #mostly for tests... FIXME: |
| 187 | } |
| 188 | simils_exo = exp(-distances2/(sd_dist*window_exo^2)) |
| 189 | } |
| 190 | |
| 191 | similarities = |
| 192 | if (simtype == "exo") |
| 193 | simils_exo |
| 194 | else if (simtype == "endo") |
| 195 | simils_endo |
| 196 | else if (simtype == "mix") |
| 197 | simils_endo * simils_exo |
| 198 | else #none |
| 199 | rep(1, length(fdays)) |
| 200 | similarities = similarities / sum(similarities) |
| 201 | |
| 202 | prediction = rep(0, horizon) |
| 203 | for (i in seq_along(fdays)) |
| 204 | prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon] |
| 205 | |
| 206 | if (final_call) |
| 207 | { |
| 208 | private$.params$weights <- similarities |
| 209 | private$.params$indices <- fdays |
| 210 | private$.params$window <- |
| 211 | if (simtype=="endo") |
| 212 | window_endo |
| 213 | else if (simtype=="exo") |
| 214 | window_exo |
| 215 | else #mix |
| 216 | c(window_endo,window_exo) |
| 217 | } |
| 218 | |
| 219 | return (prediction) |
| 220 | } |
| 221 | ) |
| 222 | ) |