fix mistake in yersteday/today computations
authorBenjamin Auder <benjamin.auder@somewhere>
Tue, 25 Apr 2017 08:06:00 +0000 (10:06 +0200)
committerBenjamin Auder <benjamin.auder@somewhere>
Tue, 25 Apr 2017 08:06:00 +0000 (10:06 +0200)
pkg/R/F_Neighbors.R
pkg/R/plot.R
pkg/R/utils.R

index ea18bb6..f140b0b 100644 (file)
@@ -45,14 +45,14 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                        private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
 
                        # Do not forecast on days with NAs (TODO: softer condition...)
-                       if (any(is.na(data$getSerie(today-1)))
-                               || any(is.na(data$getSerie(today)[1:(predict_from-1)])))
+                       if (any(is.na(data$getSerie(today-1))) ||
+                               (predict_from>=2 && any(is.na(data$getSerie(today)[1:(predict_from-1)]))))
                        {
                                return (NA)
                        }
 
-                       # Determine indices of no-NAs days followed by no-NAs tomorrows
-                       fdays = .getNoNA2(data, max(today-memory,1), today-2)
+                       # Determine indices of no-NAs days preceded by no-NAs yerstedays
+                       tdays = .getNoNA2(data, max(today-memory,2), today-1)
 
                        # Get optional args
                        local = ifelse(hasArg("local"), list(...)$local, TRUE) #same level + season?
@@ -60,12 +60,12 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                        if (hasArg("window"))
                        {
                                return ( private$.predictShapeAux(data,
-                                       fdays, today, predict_from, horizon, local, list(...)$window, simtype, TRUE) )
+                                       tdays, today, predict_from, horizon, local, list(...)$window, simtype, TRUE) )
                        }
 
                        # Indices of similar days for cross-validation; TODO: 20 = magic number
                        cv_days = getSimilarDaysIndices(today, data, limit=20, same_season=FALSE,
-                               days_in=fdays)
+                               days_in=tdays)
 
                        # Optimize h : h |--> sum of prediction errors on last N "similar" days
                        errorOnLastNdays = function(window, simtype)
@@ -75,13 +75,13 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                for (i in seq_along(cv_days))
                                {
                                        # mix_strategy is never used here (simtype != "mix"), therefore left blank
-                                       prediction = private$.predictShapeAux(data, fdays, cv_days[i], predict_from,
+                                       prediction = private$.predictShapeAux(data, tdays, cv_days[i], predict_from,
                                                horizon, local, window, simtype, FALSE)
                                        if (!is.na(prediction[1]))
                                        {
                                                nb_jours = nb_jours + 1
                                                error = error +
-                                                       mean((data$getSerie(cv_days[i]+1)[predict_from:horizon] - prediction)^2)
+                                                       mean((data$getSerie(cv_days[i])[predict_from:horizon] - prediction)^2)
                                        }
                                }
                                return (error / nb_jours)
@@ -109,41 +109,41 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                else #none: value doesn't matter
                                        1
 
-                       return( private$.predictShapeAux(data, fdays, today, predict_from, horizon, local,
+                       return( private$.predictShapeAux(data, tdays, today, predict_from, horizon, local,
                                best_window, simtype, TRUE) )
                }
        ),
        private = list(
                # Precondition: "today" is full (no NAs)
-               .predictShapeAux = function(data, fdays, today, predict_from, horizon, local, window,
+               .predictShapeAux = function(data, tdays, today, predict_from, horizon, local, window,
                        simtype, final_call)
                {
-                       fdays_cut = fdays[ fdays < today ]
-                       if (length(fdays_cut) <= 1)
+                       tdays_cut = tdays[ tdays <= today-1 ]
+                       if (length(tdays_cut) <= 1)
                                return (NA)
 
                        if (local)
                        {
                                # TODO: 60 == magic number
-                               fdays = getSimilarDaysIndices(today, data, limit=60, same_season=TRUE,
-                                       days_in=fdays_cut)
-                               if (length(fdays) <= 1)
+                               tdays = getSimilarDaysIndices(today, data, limit=60, same_season=TRUE,
+                                       days_in=tdays_cut)
+                               if (length(tdays) <= 1)
                                        return (NA)
                                # TODO: 10, 12 == magic numbers
-                               fdays = .getConstrainedNeighbs(today,data,fdays,min_neighbs=10,max_neighbs=12)
-                               if (length(fdays) == 1)
+                               tdays = .getConstrainedNeighbs(today,data,tdays,min_neighbs=10,max_neighbs=12)
+                               if (length(tdays) == 1)
                                {
                                        if (final_call)
                                        {
                                                private$.params$weights <- 1
-                                               private$.params$indices <- fdays
+                                               private$.params$indices <- tdays
                                                private$.params$window <- 1
                                        }
-                                       return ( data$getSerie(fdays[1]+1)[predict_from:horizon] )
+                                       return ( data$getSerie(tdays[1])[predict_from:horizon] )
                                }
                        }
                        else
-                               fdays = fdays_cut #no conditioning
+                               tdays = tdays_cut #no conditioning
 
                        if (simtype == "endo" || simtype == "mix")
                        {
@@ -151,9 +151,11 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                window_endo = ifelse(simtype=="mix", window[1], window)
 
                                # Distances from last observed day to days in the past
-                               lastSerie = c( data$getSerie(today-1), data$getSerie(today)[1:(predict_from-1)] )
-                               distances2 = sapply(fdays, function(i) {
-                                       delta = lastSerie - c(data$getSerie(i),data$getSerie(i+1)[1:(predict_from-1)])
+                               lastSerie = c( data$getSerie(today-1),
+                                       data$getSerie(today)[if (predict_from>=2) 1:(predict_from-1) else c()] )
+                               distances2 = sapply(tdays, function(i) {
+                                       delta = lastSerie - c(data$getSerie(i-1),
+                                               data$getSerie(i)[if (predict_from>=2) 1:(predict_from-1) else c()])
                                        sqrt(mean(delta^2))
                                })
 
@@ -165,21 +167,21 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                # Compute exogen similarities using given window
                                window_exo = ifelse(simtype=="mix", window[2], window)
 
-                               M = matrix( ncol=1+length(fdays), nrow=1+length(data$getExo(1)) )
+                               M = matrix( ncol=1+length(tdays), nrow=1+length(data$getExo(1)) )
                                M[,1] = c( data$getLevelHat(today), as.double(data$getExoHat(today)) )
-                               for (i in seq_along(fdays))
-                                       M[,i+1] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
+                               for (i in seq_along(tdays))
+                                       M[,i+1] = c( data$getLevel(tdays[i]), as.double(data$getExo(tdays[i])) )
 
                                sigma = cov(t(M)) #NOTE: robust covariance is way too slow
                                # TODO: 10 == magic number; more robust way == det, or always ginv()
                                sigma_inv =
-                                       if (length(fdays) > 10)
+                                       if (length(tdays) > 10)
                                                solve(sigma)
                                        else
                                                MASS::ginv(sigma)
 
                                # Distances from last observed day to days in the past
-                               distances2 = sapply(seq_along(fdays), function(i) {
+                               distances2 = sapply(seq_along(tdays), function(i) {
                                        delta = M[,1] - M[,i+1]
                                        delta %*% sigma_inv %*% delta
                                })
@@ -195,20 +197,20 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                else if (simtype == "mix")
                                        simils_endo * simils_exo
                                else #none
-                                       rep(1, length(fdays))
+                                       rep(1, length(tdays))
                        similarities = similarities / sum(similarities)
 
                        prediction = rep(0, horizon-predict_from+1)
-                       for (i in seq_along(fdays))
+                       for (i in seq_along(tdays))
                        {
                                prediction = prediction +
-                                       similarities[i] * data$getSerie(fdays[i]+1)[predict_from:horizon]
+                                       similarities[i] * data$getSerie(tdays[i])[predict_from:horizon]
                        }
 
                        if (final_call)
                        {
                                private$.params$weights <- similarities
-                               private$.params$indices <- fdays
+                               private$.params$indices <- tdays
                                private$.params$window <-
                                        if (simtype=="endo")
                                                window_endo
@@ -231,20 +233,20 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
 #
 # @param today Index of current day
 # @param data Object of class Data
-# @param fdays Current set of "first days" (no-NA pairs)
+# @param tdays Current set of "second days" (no-NA pairs)
 # @param min_neighbs Minimum number of points in a neighborhood
 # @param max_neighbs Maximum number of points in a neighborhood
 #
-.getConstrainedNeighbs = function(today, data, fdays, min_neighbs=10, max_neighbs=12)
+.getConstrainedNeighbs = function(today, data, tdays, min_neighbs=10, max_neighbs=12)
 {
        levelToday = data$getLevelHat(today)
        levelYersteday = data$getLevel(today-1)
-       distances = sapply(fdays, function(i) {
-               sqrt((data$getLevel(i)-levelYersteday)^2 + (data$getLevel(i+1)-levelToday)^2)
+       distances = sapply(tdays, function(i) {
+               sqrt((data$getLevel(i-1)-levelYersteday)^2 + (data$getLevel(i)-levelToday)^2)
        })
        #TODO: 1, +1, +3 : magic numbers
        dist_thresh = 1
-       min_neighbs = min(min_neighbs,length(fdays))
+       min_neighbs = min(min_neighbs,length(tdays))
        repeat
        {
                same_pollution = (distances <= dist_thresh)
@@ -253,14 +255,14 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                        break
                dist_thresh = dist_thresh + ifelse(dist_thresh>1,3,1)
        }
-       fdays = fdays[same_pollution]
+       tdays = tdays[same_pollution]
        max_neighbs = 12
        if (nb_neighbs > max_neighbs)
        {
                # Keep only max_neighbs closest neighbors
-               fdays = fdays[ order(distances[same_pollution])[1:max_neighbs] ]
+               tdays = tdays[ order(distances[same_pollution])[1:max_neighbs] ]
        }
-       fdays
+       tdays
 }
 
 # compute similarities
index 59a26a7..ad0ed4e 100644 (file)
@@ -236,10 +236,10 @@ plotRelVar = function(data, fil, predict_from)
 {
        ref_var = c( apply(data$getSeries(fil$neighb_indices),1,sd),
                apply(data$getSeries(fil$neighb_indices+1),1,sd) )
-       fdays = .getNoNA2(data, 1, fil$index-1)
+       tdays = .getNoNA2(data, 2, fil$index)
        global_var = c(
-               apply(data$getSeries(fdays),1,sd),
-               apply(data$getSeries(fdays+1),1,sd) )
+               apply(data$getSeries(tdays-1),1,sd),
+               apply(data$getSeries(tdays),1,sd) )
 
        yrange = range(ref_var, global_var)
        par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
index 96ec601..a4efd61 100644 (file)
@@ -118,7 +118,7 @@ getSimilarDaysIndices = function(index, data, limit, same_season, days_in=NULL)
 
 # getNoNA2
 #
-# Get indices in data of no-NA series followed by no-NA, within [first,last] range.
+# Get indices in data of no-NA series preceded by no-NA, within [first,last] range.
 #
 # @inheritParams dateIndexToInteger
 # @param first First index (included)
@@ -127,6 +127,6 @@ getSimilarDaysIndices = function(index, data, limit, same_season, days_in=NULL)
 .getNoNA2 = function(data, first, last)
 {
        (first:last)[ sapply(first:last, function(i)
-               !any( is.na(data$getSerie(i)) | is.na(data$getSerie(i+1)) )
+               !any( is.na(data$getSerie(i-1)) | is.na(data$getSerie(i)) )
        ) ]
 }