public void HWTTest2WeekSystemPriceOptim()
{
double[] y = { 31.0500, 30.4700, 28.9200, 27.8800, 26.9600, 27.8400, 28.7900, 28.6300, 28.4400, 28.3000, 30.6500, 31.5500, 32.1600, 32.4500, 32.6300, 33.6500, 34.9000, 36.2200, 36.6500, 36.3700, 35.4900, 34.4100, 34.6600, 32.5500, 33.1500, 32.6600, 31.8300, 31.4700, 32.5600, 34.3600, 36.2800, 38.3900, 39.0900, 38.3300, 38.4200, 38.2500, 37.9600, 37.8900, 37.8800, 38.7800, 39.8300, 39.9100, 39.3200, 38.4900, 37.4600, 36.9400, 36.3700, 34.5900, 33.1100, 32.2200, 31.4600, 31.6700, 32.0500, 33.6700, 34.9300, 35.8200, 36.3800, 36.5200, 36.7100, 36.6000, 36.5100, 36.4000, 36.4200, 36.5800, 36.9400, 36.9400, 36.8100, 36.4300, 35.9100, 35.4500, 34.7700, 32.0900, 31.7100, 30.8600, 30.2100, 30.3600, 30.8900, 32.2100, 34.3300, 35.9800, 36.7200, 36.6800, 36.8000, 36.6400, 36.4400, 36.3100, 35.8300, 36.1200, 36.8000, 37.2900, 37.1500, 36.4700, 36.0300, 35.4300, 35.1700, 32.9500, 33.6900, 32.7100, 32.0900, 32.1700, 32.7300, 33.7700, 34.0300, 33.9800, 34.4500, 36.1200, 36.6900, 36.7400, 36.5200, 36.1200, 36.1400, 36.7800, 37.7100, 38.3700, 37.9900, 36.9400, 36.2400, 35.8900, 35.7000, 35.4400, 34.8700, 33.0500, 32.2500, 32.4100, 32.4500, 32.8000, 33.2700, 32.9700, 33.6500, 34.2500, 35.0100, 35.9400, 35.7900, 35.0200, 34.8500, 35.1800, 35.9000, 37.0900, 37.6500, 37.3900, 36.9900, 36.1000, 36.0000, 34.9300, 34.9700, 33.7100, 33.1300, 33.2300, 34.1000, 35.6800, 38.5100, 40.6600, 42.4800, 42.0800, 41.9700, 41.3700, 40.3000, 39.9600, 40.1700, 41.6000, 42.3700, 41.9300, 39.9700, 38.9000, 37.9200, 37.4700, 36.4900, 35.5500 };
int period = 24;
int m = 3 * 24;
double alpha = 0.5;
double beta = 0.4;
double gamma = 0.6;
var abg = HoltWinters.OptimizeTripleHWT(y, period, m);
double[] prediction = HoltWinters.TripleHWT(y, period, m, abg.alpha, abg.beta, abg.gamma);
var tst = string.Join(" ", y.Select(x => x.ToString()));
var tstpred = string.Join(" ", prediction.Select(x => x.ToString()));
//Assert.AreEqual("Forecast does not match", expected,
// prediction, 0.0000000000001);
}
public JsonResult Forecast(DateTime?date, string forecastMethod, string spikesPreprocessMethod, double spikesThreshold,
string forecastModel, string timeHorizon, double confidence, string MathModel, string exogenousVariables)
{
var dt = date.HasValue ? date.Value : DateTime.Today;
var ths = GetUnionCaseNames <Types.TimeHorizon>();
var forecastHorizon = Array.IndexOf(ths, timeHorizon);
var forecastingMethod = GetUnionCaseFromName <ForecastMethod>(forecastMethod);
if (forecastHorizon < 0)
{
throw new ArgumentException("Wrong argument");
}
//naive hardcoded for now...
//data to estimate on is not an input parameter...
var d = AppData.GetHistoricalSeries(_timeSeries);
var data = d.Where(x => x.DateTime < dt)
.Select(x => x.Value != null ? (double)x.Value : 0) //0 for now, when interpolation is done... etc..
.ToArray();
var spikesPreprocess = GetUnionCaseFromName <SpikePreprocess>(spikesPreprocessMethod);
if (spikesPreprocess.IsSimilarDay)
{
//First deseasonalize data at larger lags, all >= weekly, a stable series is need since spikes are at lengths <= day
var desezonalizedData = Desezonalize(data, 168);
//Then perform the spike estimation
var spikeIndices = EstimateSpikesOnMovSDs(desezonalizedData, 24, 2, spikesThreshold);
//go ahead and pre process the d series... yeah...
data = ReplaceSingularSpikes(data, spikeIndices, spikesPreprocess, 0.95);
}
var horizon = GetTimeHorizonValue(forecastHorizon);
var rlzd = d.Where(x => x.DateTime >= dt)
.Take(horizon)
.Select(x => x.Value != null ? (double)x.Value : 0) //0 for now, when interpolation is done... etc..
.ToArray();
ForecastResult forecast;
double[] forecasted;
object model;
var oneYear = (int)Math.Floor((dt - dt.AddMonths(-12)).TotalHours);
var halfYear = (int)Math.Floor((dt - dt.AddMonths(-6)).TotalHours);
var seasons = new int[] { oneYear, halfYear, 24 * 7 * 4 * 3, 24 * 7 * 4, 24 * 7, 24 };
//we need always seasonalities >= forecast horizon
var relevantSeasons = seasons.Where(s => s >= horizon).ToArray();
double[] estimationData = null;
//add naive with exogenous variables next...
if (forecastingMethod == ForecastMethod.Naive)
{
forecast = Naive(data, relevantSeasons, horizon, confidence);
forecasted = forecast.Forecast.Take(rlzd.Length).ToArray();
model = new object();
}
else if (forecastingMethod == ForecastMethod.HoltWinters)
{
estimationData = data.Reverse().Take(24 * 7 * 2).Reverse().ToArray();
var hwparams = JsonConvert.DeserializeObject <HoltWinters.HoltWintersParams>(MathModel);
//optimize if all seem to be wrong
if (hwparams.alpha == 0 || hwparams.beta == 0 || hwparams.gamma == 0)
{
hwparams = HoltWinters.OptimizeTripleHWT(estimationData, 24, horizon);
}
forecast = HoltWinters.TripleHWTWithPIs(estimationData, 24, horizon, hwparams, confidence);
forecasted = forecast.Forecast.Take(rlzd.Length).ToArray();
model = hwparams;
}
else if (forecastingMethod == ForecastMethod.ARMA)
{
var exogenousVariablesJs = JObject.Parse(exogenousVariables); //alternative...?
//always take next 2 highest seasons...
var maSes = seasons.Where(s => s >= horizon).OrderBy(x => x).Take(2).ToArray();
var arSes = new int[] { 1, 2, 24, 25 };
//twice as the highest season estimation data
var estimationDataLength = maSes.Last() * 2;
estimationData = data.Reverse().Take(estimationDataLength).Reverse().ToArray();
var isUnivariate = exogenousVariablesJs.Properties().Where(p => p.Value.ToString() == "True").Count() == 0;
if (isUnivariate)
{
var arma = JsonConvert.DeserializeObject <ARMAResult>(MathModel);
//problem: when changing param for ARMA, provide forecast, when changed date or horizon, re-estimate...yes...
//fix it later
if (arma.AR == null || arma.AR.Coefficients == null)
{
}
arma = ARMASimple2(estimationData, arSes, maSes);
var inSampleRes = Infer(arma, estimationData);
forecast = MarketModels.TimeSeries.Forecast(estimationData, inSampleRes, arma, horizon, confidence);
//when done out of sample, matches the realized and forecasted lengths to compute fit...
//if no available data, no fit can be done...
forecasted = forecast.Forecast.Take(rlzd.Length).ToArray();
//allow only AR and MA coefficients to be changed
//model = new { AR = arma.AR, MA = arma.MA };
model = null;
}
else
{
var arxma = JsonConvert.DeserializeObject <ARXMAModel>(MathModel);
var vars = exogenousVariablesJs.Properties()
.Where(p => p.Value.ToString() == "True")
.Select(p => p.Name)
.ToList();
var specialDaysSelected = vars.Remove("SpecialWeekDays");
var exData = vars
.Select(p =>
AppData.GetHistoricalSeries(_exVarsPre + p + "_Hourly_All.json")
.Where(x => x.DateTime < dt)
.Select(x => x.Value != null ? (double)x.Value : 0) //0 for now, when interpolation is done... etc..
.Reverse().Take(estimationDataLength + horizon).Reverse().ToArray()
.ToList()
).ToList().ColumnsToRectangularArray();
//append indicators for special data...
if (specialDaysSelected)
{
//Friday, Saturday and Sunday
var ivars = MathFunctions.IndicatorVariablesMatrix(estimationDataLength + horizon, 168, 24, new int[] { 0, 5, 6 });
//concatenate matrices to columns
if (exData.GetLength(1) > 0)
{
exData = MathFunctions.concat2D(exData, ivars, false);
}
else
{
exData = ivars;
}
}
var exDataEst = MathFunctions.firstRows2D(exData, estimationDataLength);
arxma = ARXMASimple2(estimationData, exDataEst, arSes, maSes);
var inSampleRes = arxma.Infer(estimationData, exDataEst);
forecast = arxma.Forecast(estimationData, inSampleRes, exData, horizon, confidence);
//when done out of sample, matches the realized and forecasted lengths to compute fit...
//if no available data, no fit can be done...
forecasted = forecast.Forecast.Take(rlzd.Length).ToArray();
//allow only AR and MA coefficients to be changed
//model = new { AR = arma.AR, MA = arma.MA };
model = null;
}
}
else
{
throw new ArgumentException("Unsupported argument passed");
}
var log = "";
if (forecast.Forecast.Any(x => x > 500))
{
log += "Some data is wrong...";
}
if (forecast.Forecast.HasInvalidData() || forecast.Confidence.HasInvalidData())
{
throw new Exception("Abnormal results generated");
}
/* Post processing */
//cap all values above a reasonable limit, e.g. 500
CapSeries(forecast.Forecast, 500, -100);
CapMultipleSeries(forecast.Confidence, 500, -100);
FitStatistics eFit = null, eBFit = null, ePFit = null, fit = null, bfit = null, pfit = null;
if (rlzd.Length > 0)
{
fit = ForecastFit(forecasted, rlzd);
pfit = ForecastFit(
GetSubPeriodsFrom(forecasted, 24, DAY_PEAK_HOURS),
GetSubPeriodsFrom(rlzd, 24, DAY_PEAK_HOURS));
bfit = ForecastFit(
GetSubPeriodsFrom(forecasted, 24, DAY_BASE_HOURS),
GetSubPeriodsFrom(rlzd, 24, DAY_BASE_HOURS));
}
if (estimationData != null && forecast.Backcast.Length > 0)
{
//model with estimation, that is: the model is different than naive
eFit = ForecastFit(forecast.Backcast, estimationData);
ePFit = ForecastFit(
GetSubPeriodsFrom(forecast.Backcast, 24, DAY_PEAK_HOURS),
GetSubPeriodsFrom(estimationData, 24, DAY_PEAK_HOURS));
eBFit = ForecastFit(
GetSubPeriodsFrom(forecast.Backcast, 24, DAY_BASE_HOURS),
GetSubPeriodsFrom(estimationData, 24, DAY_BASE_HOURS));
}
var obj = new
{
Result = forecast,
MathModel = model,
DaysAhead = horizon / 24,
//in sample fit for all hours
EstimationFit = eFit,
BaseEstimationFit = eBFit,
PeakEstimationFit = ePFit,
Fit = fit,
BaseFit = bfit, //refine later...
PeakFit = pfit,
//additional remarks about the resulting data...
Log = log
};
return(Json(obj, JsonRequestBehavior.AllowGet));
}