Repository Public
1import Foundation23extension Inference {4 /// Jarque–Bera test of normality: JB = n/6·(S² + K²/4) with the population skewness S and excess5 /// kurtosis K; asymptotically χ²(2), upper-tail p-value. Needs large samples (the χ² approximation is poor for n < ~50).6 public static func jarqueBera(_ x: [Double]) throws -> HypothesisTest {7 let v = try clean(x, min: 3, "sample")8 let s = Descriptive.skewness(v, sample: false), k = Descriptive.kurtosis(v, sample: false)9 guard s.isFinite, k.isFinite else { throw StatError.invalidInput("the sample is constant") }10 let jb = Double(v.count) / 6 * (s * s + k * k / 4)11 return HypothesisTest(name: "Jarque–Bera normality test", statisticName: "JB", statistic: jb, df: 2,12 pValue: ChiSquaredDistribution(df: 2).sf(jb), alternative: .greater, pValueDescription: "upper tail of χ²(2)",13 details: [NamedValue("n", Double(v.count)), NamedValue("Skewness (population)", s), NamedValue("Excess kurtosis (population)", k)],14 assumptions: ["Observations are independent and identically distributed", "Large sample (asymptotic χ² reference distribution)"])15 }1617 private static func poly(_ c: [Double], _ x: Double) -> Double {18 var r = 0.019 for coef in c.reversed() { r = r * x + coef }20 return r21 }2223 /// Shapiro–Wilk test of normality by Royston's algorithm AS R94 (1995), valid for 3 ≤ n ≤ 5000.24 /// W close to 1 supports normality; the p-value is the probability of a W this small or smaller25 /// under normality (Royston's normalising transformation; exact for n = 3).26 public static func shapiroWilk(_ x: [Double]) throws -> HypothesisTest {27 let v = try clean(x, min: 3, "sample").sorted()28 let n = v.count29 guard n <= 5000 else { throw StatError.invalidInput("Royston's approximation is only valid up to n = 5000") }30 let range = v[n - 1] - v[0]31 guard range > 0 else { throw StatError.invalidInput("the sample is constant") }32 let nn2 = n / 233 var a = [Double](repeating: 0, count: nn2) // weights of the pairs (x(n+1−i) − x(i))34 if n == 3 {35 a[0] = 0.5.squareRoot()36 } else {37 let an = Double(n)38 var m = [Double](repeating: 0, count: nn2)39 var summ2 = 0.040 for i in 0..<nn2 {41 m[i] = SpecialFunctions.normalQuantile((Double(i + 1) - 0.375) / (an + 0.25))42 summ2 += m[i] * m[i]43 }44 summ2 *= 245 let ssumm2 = summ2.squareRoot()46 let rsn = 1 / an.squareRoot()47 let a1 = poly([0, 0.221157, -0.147981, -2.071190, 4.434685, -2.706056], rsn) - m[0] / ssumm248 var first: Int, fac: Double49 if n > 5 {50 let a2 = -m[1] / ssumm2 + poly([0, 0.042981, -0.293762, -1.752461, 5.682633, -3.582633], rsn)51 fac = ((summ2 - 2 * m[0] * m[0] - 2 * m[1] * m[1]) / (1 - 2 * a1 * a1 - 2 * a2 * a2)).squareRoot()52 a[0] = a1; a[1] = a253 first = 254 } else {55 fac = ((summ2 - 2 * m[0] * m[0]) / (1 - 2 * a1 * a1)).squareRoot()56 a[0] = a157 first = 158 }59 if first < nn2 { for i in first..<nn2 { a[i] = -m[i] / fac } }60 }61 var b = 0.062 for i in 0..<nn2 { b += a[i] * (v[n - 1 - i] - v[i]) }63 let ss = Descriptive.sumSquaredDeviations(v)64 let w = min(1, b * b / ss)6566 var p: Double67 if n == 3 {68 p = max(0, 6 / Double.pi * (asin(w.squareRoot()) - asin(0.75.squareRoot())))69 } else {70 let an = Double(n)71 let y = log(1 - w)72 if n <= 11 {73 let gamma = poly([-2.273, 0.459], an)74 if y >= gamma {75 p = 076 } else {77 let yy = -log(gamma - y)78 let mm = poly([0.5440, -0.39978, 0.025054, -6.714e-4], an)79 let s = exp(poly([1.3822, -0.77857, 0.062767, -0.0020322], an))80 p = NormalDistribution.standard.sf((yy - mm) / s)81 }82 } else {83 let xx = log(an)84 let mm = poly([-1.5861, -0.31082, -0.083751, 0.0038915], xx)85 let s = exp(poly([-0.4803, -0.082676, 0.0030302], xx))86 p = NormalDistribution.standard.sf((y - mm) / s)87 }88 }89 if w == 1 { p = 1 }90 return HypothesisTest(name: "Shapiro–Wilk normality test", statisticName: "W", statistic: w, pValue: p, alternative: .less,91 pValueDescription: "probability of W this small or smaller under normality (Royston 1995)",92 details: [NamedValue("n", Double(n))],93 assumptions: ["Observations are independent and identically distributed", "3 ≤ n ≤ 5000; heavy ties (rounded data) reduce W"])94 }95}96