Improvements and bugfixes on pwfts.forecast_interval and pwfts.point_expected_value

pyFTS/benchmarks/benchmarks.py

@@ -16,6 +16,7 @@ import numpy as np
 from mpl_toolkits.mplot3d import Axes3D
 
 from pyFTS.probabilistic import ProbabilityDistribution
+from pyFTS.common import Transformations
 from pyFTS.models import song, chen, yu, ismailefendi, sadaei, hofts, pwfts, ifts, cheng, hwang
 from pyFTS.models.ensemble import ensemble
 from pyFTS.benchmarks import Measures, naive, arima, ResidualAnalysis, quantreg
@@ -833,10 +834,6 @@ def plot_density_rectange(ax, cmap, density, fig, resolution, time_from, time_to
     cb.set_label('Density')
 
 
-from pyFTS.common import Transformations
-
-
-
 def plot_distribution(ax, cmap, probabilitydist, fig, time_from, reference_data=None):
     from matplotlib.patches import Rectangle
     from matplotlib.collections import PatchCollection

pyFTS/models/pwfts.py

@@ -257,6 +257,16 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
         return ret
 
     def forecast(self, data, **kwargs):
+        method = kwargs.get('method','heuristic')
+
+        if method == 'heuristic':
+            return self.point_heuristic(data, **kwargs)
+        elif method == 'expected_value':
+            return self.point_expected_value(data, **kwargs)
+        else:
+            raise Exception("Unknown point forecasting method!")
+
+    def point_heuristic(self, data, **kwargs):
 
         ndata = np.array(self.apply_transformations(data))
 
@@ -292,10 +302,26 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
 
         return ret
 
+    def point_expected_value(self, data, **kwargs):
+        l = len(data)
+
+        ret = []
+
+        for k in np.arange(self.order - 1, l):
+            sample = data[k - (self.order - 1): k + 1]
+
+            tmp = self.forecast_distribution(sample)[0].expected_value()
+
+            ret.append(tmp)
+
+        ret = self.apply_inverse_transformations(ret, params=[data[self.order - 1:]])
+
+        return ret
+
     def forecast_interval(self, data, **kwargs):
 
         if 'method' in kwargs:
-            self.interval_method = kwargs.get('method','quantile')
+            self.interval_method = kwargs.get('method','heuristic')
 
         if 'alpha' in kwargs:
             self.alpha = kwargs.get('alpha', 0.05)
@@ -308,8 +334,8 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
 
         for k in np.arange(self.order - 1, l):
 
-            if self.interval_method == 'extremum':
+            if self.interval_method == 'heuristic':
-                self.interval_extremum(k, ndata, ret)
+                self.interval_heuristic(k, ndata, ret)
             else:
                 self.interval_quantile(k, ndata, ret)
 
@@ -321,9 +347,9 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
         dist = self.forecast_distribution(ndata)
         lo_qt = dist[0].quantile(self.alpha)
         up_qt = dist[0].quantile(1.0 - self.alpha)
-        ret.append_rhs([lo_qt, up_qt])
+        ret.append([lo_qt, up_qt])
 
-    def interval_extremum(self, k, ndata, ret):
+    def interval_heuristic(self, k, ndata, ret):
 
         sample = ndata[k - (self.order - 1): k + 1]
 
@@ -335,7 +361,7 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
         for flrg in flrgs:
             norm = self.flrg_lhs_conditional_probability(sample, flrg)
             if norm == 0:
-                norm = self.flrg_lhs_unconditional_probability(flrg)  # * 0.001
+                norm = self.flrg_lhs_unconditional_probability(flrg)
             up.append(norm * self.get_upper(flrg))
             lo.append(norm * self.get_lower(flrg))
             norms.append(norm)
@@ -349,7 +375,6 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
             up_ = sum(up) / norm
             ret.append([lo_, up_])
 
-
     def forecast_distribution(self, data, **kwargs):
 
         if not isinstance(data, (list, set, np.ndarray)):
@@ -425,8 +450,8 @@ class ProbabilisticWeightedFTS(ifts.IntervalFTS):
                 1] >= self.original_max):
                 ret.append(ret[-1])
             else:
-                lower = self.forecast_interval([ret[x][0] for x in np.arange(k - self.order, k)])
+                lower = self.forecast_interval([ret[x][0] for x in np.arange(k - self.order, k)], **kwargs)
-                upper = self.forecast_interval([ret[x][1] for x in np.arange(k - self.order, k)])
+                upper = self.forecast_interval([ret[x][1] for x in np.arange(k - self.order, k)], **kwargs)
 
                 ret.append([np.min(lower), np.max(upper)])
 

pyFTS/probabilistic/ProbabilityDistribution.py

@@ -95,6 +95,9 @@ class ProbabilityDistribution(object):
 
         return ret
 
+    def expected_value(self):
+        return np.nansum([v * self.distribution[v] for v in self.bins])
+
     def build_cdf_qtl(self):
         ret = 0.0
         self.cdf = {}
@@ -137,7 +140,7 @@ class ProbabilityDistribution(object):
                 ret.append(self.qtl[str(k)][0])
         else:
             k = self.quantile_index.find_ge(values)
-            ret = self.qtl[str(k)[0]]
+            ret = self.qtl[str(k)]
 
         return ret
 

pyFTS/tests/general.py

@@ -9,6 +9,8 @@ import matplotlib.pylab as plt
 import pandas as pd
 from pyFTS.common import Transformations
 
+tdiff = Transformations.Differential(1)
+
 from pyFTS.data import TAIEX
 
 dataset = TAIEX.get_data()
@@ -30,17 +32,28 @@ test_length = 200
 
 from pyFTS.partitioners import Grid, Util as pUtil
 partitioner = Grid.GridPartitioner(data=dataset[:train_split], npart=30)
+#partitioner = Grid.GridPartitioner(data=dataset[:train_split], npart=10, transformation=tdiff)
 
 from pyFTS.common import fts,tree
 from pyFTS.models import hofts, pwfts
 
 pfts1_taiex = pwfts.ProbabilisticWeightedFTS("1", partitioner=partitioner)
+#pfts1_taiex.append_transformation(tdiff)
 pfts1_taiex.fit(dataset[:train_split], save_model=True, file_path='pwfts')
 pfts1_taiex.shortname = "1st Order"
 
 print(pfts1_taiex)
 
-tmp = pfts1_taiex.predict(dataset[train_split:train_split+200], type='distribution', steps_ahead=20)
+tmp = pfts1_taiex.predict(dataset[train_split:train_split+200], type='interval',
+                          method='quantile', alpha=.05, steps_ahead=10)
+
+
+
+'''
+tmp = pfts1_taiex.predict(dataset[train_split:train_split+200], type='diPedro Pazzini
+stribution', steps_ahead=20)
+
+
 f, ax = plt.subplots(3, 4, figsize=[20,15])
 tmp[0].plot(ax[0][0], title='t=1')
 tmp[2].plot(ax[0][1], title='t=20')
@@ -52,4 +65,10 @@ tmp[12].plot(ax[1][2], title='t=120')
 tmp[14].plot(ax[1][3], title='t=140')
 tmp[16].plot(ax[2][0], title='t=160')
 tmp[18].plot(ax[2][1], title='t=180')
-tmp[20].plot(ax[2][2], title='t=200')
+tmp[20].plot(ax[2][2], title='t=200')
+
+
+f, ax = plt.subplots(1, 1, figsize=[20,15])
+bchmk.plot_distribution(ax, 'blue', tmp, f, 0, reference_data=dataset[train_split:train_split+200])
+
+'''