// an extra randomness reduction was helpful; see http://neuralsniffer.wordpress.com/2010/12/13/combining-the-3-inputs-1-black-box-or-3-black-boxes-with-confidence/
public double GetEnsembleRepeatForecast(int p_nEnsembleRepeat, EnsembleGroupSetup[] p_ensembleGroups, EnsembleAggregationStrategy p_ensembleAggregation, int p_maxEpoch, int p_iRebalance, int p_lookbackWindowSize, double p_outputOutlierThreshold, double p_inputOutlierClipInSD, double p_inputNormalizationBoost, double p_outputNormalizationBoost, int p_notNNStrategyType, double[] p_dateWeekDays, double[] p_barChanges, bool p_isLogForecast, out double p_avgTrainError)
{
double[] forecasts = new double[p_nEnsembleRepeat];
double[] trainErrors = new double[p_nEnsembleRepeat];
for (int i = 0; i < forecasts.Length; i++)
{
double trainError = Double.NaN;
forecasts[i] = GetEnsembleForecast(p_ensembleGroups, p_ensembleAggregation, p_maxEpoch, p_iRebalance, p_lookbackWindowSize, p_outputOutlierThreshold, p_inputOutlierClipInSD, p_inputNormalizationBoost, p_outputNormalizationBoost, p_notNNStrategyType, p_dateWeekDays, p_barChanges, p_isLogForecast, out trainError);
trainErrors[i] = trainError;
}
p_avgTrainError = trainErrors.Average();
if (p_isLogForecast)
Utils.Logger.Info("Forecast%: GetEnsembleRepeatForecast(): " + String.Concat(forecasts.Select(r => (r * 100).ToString("F4") + "%,").ToArray()));
return forecasts.Select(r => Math.Sign(r)).Sum();
}
double GetEnsembleForecast(EnsembleGroupSetup[] p_ensembleGroups, EnsembleAggregationStrategy p_ensembleAggregation, int p_maxEpoch, int p_iRebalance, int p_lookbackWindowSize, double p_outputOutlierThreshold, double p_inputOutlierClipInSD, double p_inputNormalizationBoost, double p_outputNormalizationBoost, int p_notNNStrategyType, double[] p_dateWeekDays, double[] p_barChanges, bool p_isLogForecast, out double p_avgTrainError)
{
List<double>[] forecasts = new List<double>[p_ensembleGroups.Length];
List<double>[] trainErrors = new List<double>[p_ensembleGroups.Length];
for (int i = 0; i < p_ensembleGroups.Length; i++)
{
EnsembleGroupSetup group = p_ensembleGroups[i];
forecasts[i] = new List<double>();
trainErrors[i] = new List<double>();
//PrepareMemberForecast(); TODO: prepare input/output only once
int nMembersInGroup = group.NensembleGroupMembers;
for (int m = 0; m < nMembersInGroup; m++)
{
double trainError = Double.NaN;
forecasts[i].Add(GetMemberForecast(group.NNInputDesc, group.Nneurons, p_maxEpoch, p_iRebalance, p_lookbackWindowSize, group.BarChangeLookbackDaysInds, p_outputOutlierThreshold, p_inputOutlierClipInSD, p_inputNormalizationBoost, p_outputNormalizationBoost, p_notNNStrategyType, p_dateWeekDays, p_barChanges, out trainError));
trainErrors[i].Add(trainError);
}
if (p_isLogForecast)
Utils.Logger.Info("Forecast%: GetEnsembleForecast(): " + String.Concat(forecasts[i].Select(r => (r * 100).ToString("F4") + "%,").ToArray()));
}
int nTotalMember = p_ensembleGroups.Select(r => r.NensembleGroupMembers).Sum();
StrongAssert.True(nTotalMember > 0 && p_ensembleGroups.Length != 0);
p_avgTrainError = trainErrors.Select(r => r.Sum()).Sum() / (double)nTotalMember;
switch (p_ensembleAggregation) // instead of returning Sum(), return SIGN(Sum()), because the Meta mechanism will repeat this thing again: OK; we don't need; the Meta will take the Sign() at first
{
case EnsembleAggregationStrategy.ReturnTheFirstForecast:
return forecasts[0][0];
case EnsembleAggregationStrategy.SumSignForecasts:
return forecasts.Select(r => r.Select(p => Math.Sign(p)).Sum()).Sum(); // SumSign() has lowest STD in the end than Avg() (I measured it)
case EnsembleAggregationStrategy.AvgForecasts:
return forecasts.Select(r => r.Sum()).Sum() / (double)nTotalMember;
case EnsembleAggregationStrategy.PlayOnlyIfAllMemberAgree:
{
int nPosForecasts = forecasts.Select(r => r.Count(p => p > 0)).Sum();
int nNegForecasts = forecasts.Select(r => r.Count(p => p < 0)).Sum();
StrongAssert.True(nPosForecasts + nNegForecasts <= nTotalMember);
if (nPosForecasts == nTotalMember || nNegForecasts == nTotalMember) // we asserted that nTotalMember != 0
return forecasts.Select(r => r.Sum()).Sum() / (double)nTotalMember; // % average is OK, because they are the same sign
else
return 0; // if sign differs, go to cash
}
case EnsembleAggregationStrategy.PlayOnlyIfAllGroupAgree_onGroupSumSign: // if sign differs, go to cash
{
int nPosGroupForecasts = forecasts.Select(r => r.Select(p => Math.Sign(p)).Sum()).Count(k => k > 0);
int nNegGroupForecasts = forecasts.Select(r => r.Select(p => Math.Sign(p)).Sum()).Count(k => k < 0);
StrongAssert.True(nPosGroupForecasts + nNegGroupForecasts <= p_ensembleGroups.Length);
if (nPosGroupForecasts == p_ensembleGroups.Length || nNegGroupForecasts == p_ensembleGroups.Length) // we asserted that p_ensembleGroups.Length != 0
return forecasts[0].Select(r => Math.Sign(r)).Sum(); // % the groups are the same sign. Return the first group sign; no point averaging the SumSign values; if 5 items gives negative, it will return '-5'
else
return 0; // if sign differs, go to cash
}
default:
throw new NotImplementedException();
}
throw new NotImplementedException();
}
