Warrior_EA/System/BinomialStats.mqh

//+------------------------------------------------------------------+
//|                                                  BinomialStats.mqh |
//|                                                      AnimateDread  |
//| project. Free functions, no state, so the deploy gate, the two   |
//| edge floors, the barrier ladder and the detectability reports    |
//| all read the same formula instead of nine transcriptions of it.  |
//+------------------------------------------------------------------+
#ifndef WARRIOR_SYSTEM_BINOMIALSTATS_MQH
#define WARRIOR_SYSTEM_BINOMIALSTATS_MQH
#include <Math\Stat\Normal.mqh>
//+------------------------------------------------------------------+
//| Upper-tail standard normal, Q(z) = P(Z >= z).                     |
//|                                                                   |
//| Wraps the library so the NaN policy lives in one place: an        |
//| unusable z reads as "not significant" rather than propagating a   |
//| NaN into a gate decision. tail=false asks for the UPPER tail, and |
//| the clamp keeps a -1e-17 round-off out of the Sidak power.        |
//+------------------------------------------------------------------+
double NormalUpperTailQ(const double z)
  {
   if(!MathIsValidNumber(z))
      return(1.0);
   int    err=0;
   double q=MathCumulativeDistributionNormal(z,0.0,1.0,false,false,err);
   if(err!=ERR_OK || !MathIsValidNumber(q))
      return(1.0);
   return(MathMax(0.0,MathMin(1.0,q)));
  }
//+------------------------------------------------------------------+
//| Sampling variance of a binomial rate p over n observations:       |
//| p(1-p)/n, in fraction^2.                                          |
//|                                                                   |
//| Returns 0 for a degenerate rate or an empty sample, which every   |
//| caller already treats as "no bar to clear". Callers that combine  |
//| symbols by inverse variance want this rather than the SE.         |
//+------------------------------------------------------------------+
double BinomialVar(const double p,const double n)
  {
   if(!MathIsValidNumber(p) || !MathIsValidNumber(n))
      return(0.0);
   if(n<=0.0 || p<=0.0 || p>=1.0)
      return(0.0);
   return(p*(1.0-p)/n);
  }
//+------------------------------------------------------------------+
//| Standard error of a binomial rate p over n observations, in       |
//| percentage points: 100 * sqrt(p(1-p)/n).                          |
//+------------------------------------------------------------------+
double BinomialSEPct(const double p,const double n)
  {
   return(100.0*MathSqrt(BinomialVar(p,n)));
  }
//+------------------------------------------------------------------+
//| Independent observations needed to certify an edge of `edge` over |
//| a base rate p, at `sigmas` standard errors:                       |
//|   n = sigmas^2 * p(1-p) / edge^2                                  |
//|                                                                   |
//| BinomialSEPct solved for n. Answers "could this configuration     |
//| EVER prove an edge this size" - a property of the geometry, the   |
//| horizon and the window, which no amount of training moves.        |
//+------------------------------------------------------------------+
double BinomialCallsForEdge(const double p,const double edge,const double sigmas)
  {
   if(edge<=0.0 || p<=0.0 || p>=1.0)
      return(0.0);
   return(sigmas*sigmas*p*(1.0-p)/(edge*edge));
  }
//+------------------------------------------------------------------+
//| A rate shrunk toward a prior - the estimator, where everything   |
//| above is the test.                                               |
//+------------------------------------------------------------------+
double ShrunkRatePct(const double hits,const double n,const double priorPct,const double priorN)
  {
   bool havePrior=(priorN>0.0 && MathIsValidNumber(priorPct) && priorPct>=0.0);
   if(!MathIsValidNumber(hits) || !MathIsValidNumber(n) || n<=0.0)
      return(havePrior ? priorPct : 0.0);   // no evidence => the prior IS the estimate
   if(!havePrior)
      return(100.0*hits/n);
   return((hits+priorN*(priorPct/100.0))*100.0/(n+priorN));
  }
//+------------------------------------------------------------------+
//| Sidak family-wise p for the best of nTried candidates:            |
//| 1 - (1 - p_single)^N.                                             |
//|                                                                   |
//| The null of the MAXIMUM, not of a single draw. At the magnitudes  |
//| in play (p ~ 1e-4..1e-2, N ~ 10..1000) plain double precision is  |
//| ample - no need for the log1p/expm1 form MQL5 would not give us   |
//| anyway.                                                           |
//+------------------------------------------------------------------+
double SidakFamilyP(const double zObs,const int nTried)
  {
   return(1.0-MathPow(1.0-NormalUpperTailQ(zObs),(double)MathMax(nTried,1)));
  }
//+------------------------------------------------------------------+
#endif