[{"data":1,"prerenderedAt":863},["ShallowReactive",2],{"article-alternates":3,"article-\u002Fde\u002Fmarketing\u002Fbayesian-ab-test-schnelle-entscheidungen":12},{"i18nKey":4,"paths":5},"marketing-002-2026-05",{"de":6,"en":7,"es":8,"fr":9,"it":10,"ru":11},"\u002Fde\u002Fmarketing\u002Fbayesian-ab-test-schnelle-entscheidungen","\u002Fen\u002Fmarketing\u002Ffast-decisions-with-bayesian-ab-testing","\u002Fes\u002Fmarketing\u002Fbayesian-ab-test-hizli-karar","\u002Ffr\u002Fmarketing\u002Ftest-karar-bayesian","\u002Fit\u002Fmarketing\u002Ftest-bayesian-decisione-rapida","\u002Fru\u002Fmarketing\u002Fbayesian-ab-testy-dlya-bystrogo-prinyatiya-reshenij",{"_path":6,"_dir":13,"_draft":14,"_partial":14,"_locale":15,"title":16,"description":17,"publishedAt":18,"modifiedAt":18,"category":13,"i18nKey":4,"tags":19,"readingTime":25,"author":26,"body":27,"_type":857,"_id":858,"_source":859,"_file":860,"_stem":861,"_extension":862},"marketing",false,"","Bayesian A\u002FB-Test für schnellere Entscheidungsfindung","Überwinden Sie die Sample-Size-Falle von Frequentist-Tests. Der Bayesian-Ansatz mit sequentieller Überwachung und vorzeitiger Beendigung verkürzt Testprozesse um 40–60 %.","2026-05-30",[20,21,22,23,24],"ab-testing","bayesian-statistics","experimentation","conversion-optimization","statistical-inference",9,"Roibase",{"type":28,"children":29,"toc":848},"root",[30,38,45,50,61,71,81,97,103,108,118,128,147,157,167,172,179,663,669,674,684,701,711,716,721,727,732,742,752,757,767,777,782,788,793,803,813,823,833,837,842],{"type":31,"tag":32,"props":33,"children":34},"element","p",{},[35],{"type":36,"value":37},"text","Im Performance-Marketing ist der A\u002FB-Test das Rückgrat der evidenzgestützten Entscheidungsfindung. Doch viele Teams bleiben in der Dogmatik der frequentistischen Statistik stecken – dem Dogma der fixen Stichprobengröße: „Schaue nicht, bevor du die berechnete Zahl erreichst, oder du erzeugst Bias.\" Dieser Ansatz zieht Testprozesse unnötigerweise auf 3–4 Wochen in die Länge. Der Bayesian A\u002FB-Test erlaubt sequentielle Überwachung mit Posterior-Wahrscheinlichkeit. Du liest Daten täglich, kombinierst sie mit Vorwissen und beendest den Test, sobald du einen Vertrauensschwellwert erreichst (z. B. 95 % Wahrscheinlichkeit, der beste zu sein). Resultat: Die gleiche statistische Zuverlässigkeit, aber 40–60 % schneller entscheiden.",{"type":31,"tag":39,"props":40,"children":42},"h2",{"id":41},"strukturelle-grenzen-des-frequentistischen-ansatzes",[43],{"type":36,"value":44},"Strukturelle Grenzen des frequentistischen Ansatzes",{"type":31,"tag":32,"props":46,"children":47},{},[48],{"type":36,"value":49},"Der frequentistische A\u002FB-Test ruht auf p-Wert und Konfidenzintervall. Du testest die Nullhypothesen-Signifikanz – versuchst, die Annahme „es gibt keinen Unterschied zwischen Variante A und B\" abzulehnen. Die Kernprobleme dieses Ansatzes:",{"type":31,"tag":32,"props":51,"children":52},{},[53,59],{"type":31,"tag":54,"props":55,"children":56},"strong",{},[57],{"type":36,"value":58},"Verpflichtung auf feste Stichprobengröße.",{"type":36,"value":60}," Du führst eine Power-Analyse durch: Baseline-Konversionsrate 2 %, minimale nachweisbare Effektgröße (MDE) 10 % relativer Lift, Alpha 0,05, Power 0,80. Die berechnete Stichprobengröße (z. B. 15.000 Impressionen pro Variante) ist verpflichtend zu erreichen. Wenn du früh schaust und stoppen möchtest, tritt das Multiple-Comparison-Problem auf – die False-Positive-Rate übersteigt Alpha (0,05). In der Praxis: Du siehst am 2. Tag 25 % Lift, wartest aber weitere 3 Wochen, weil „die Daten nicht ausreichen.\"",{"type":31,"tag":32,"props":62,"children":63},{},[64,69],{"type":31,"tag":54,"props":65,"children":66},{},[67],{"type":36,"value":68},"Unzulängliche Ausdruckskraft von Posterior-Unsicherheit.",{"type":36,"value":70}," Der p-Wert sagt dir „die Wahrscheinlichkeit, dieses oder ein extremeres Ergebnis unter der Nullhypothese zu sehen.\" Aber was du wirklich brauchst: „Wie wahrscheinlich ist es, dass Variante B wirklich besser ist?\" Der frequentistische Rahmen antwortet auf diese Frage nicht direkt – p \u003C 0,05 ist nur die Schwelle zur Nullhypothesen-Ablehnung, nicht ein Maß für B's Überlegenheit.",{"type":31,"tag":32,"props":72,"children":73},{},[74,79],{"type":31,"tag":54,"props":75,"children":76},{},[77],{"type":36,"value":78},"Binärer Entscheidungsmechanismus.",{"type":36,"value":80}," Ist der p-Wert 0,049, ist es „signifikant\"; ist er 0,051, ist es „nicht signifikant\". Die echte Welt ist nicht so scharf. Du kannst einen p-Wert von 0,06 nicht als „marginale Evidenz, aber Test sollte verlängert werden\" interpretieren – es bleibt ein Ja oder Nein.",{"type":31,"tag":32,"props":82,"children":83},{},[84,86,95],{"type":36,"value":85},"Diese strukturellen Grenzen drücken die Test-Velocity herunter, besonders in Prozessen der ",{"type":31,"tag":87,"props":88,"children":92},"a",{"href":89,"rel":90},"https:\u002F\u002Fwww.roibase.com.tr\u002Fde\u002Fcro",[91],"nofollow",[93],{"type":36,"value":94},"Conversion Rate Optimization",{"type":36,"value":96},". Statt 2–3 Hypothesen-Iterationen pro Woche zu drehen, bleibst du an Stichprobengrößen-Regeln stecken.",{"type":31,"tag":39,"props":98,"children":100},{"id":99},"bayesian-test-posterior-wahrscheinlichkeit-und-sequentielle-überwachung",[101],{"type":36,"value":102},"Bayesian-Test: Posterior-Wahrscheinlichkeit und sequentielle Überwachung",{"type":31,"tag":32,"props":104,"children":105},{},[106],{"type":36,"value":107},"Der Bayesian-Ansatz behandelt den Parameter (Konversionsrate) nicht als feste Zahl, sondern als Wahrscheinlichkeitsverteilung. Prior-Überzeugung (Vorwissen) + beobachtete Daten → Posterior-Verteilung (aktualisierter Glaube). Das mathematische Fundament:",{"type":31,"tag":32,"props":109,"children":110},{},[111,116],{"type":31,"tag":54,"props":112,"children":113},{},[114],{"type":36,"value":115},"Prior-Verteilung:",{"type":36,"value":117}," Dein Vorwissen über die Baseline-Konversionsrate. Ohne Wissen nutzt du einen uninformierten Prior (Beta(1,1)) – gleiche Wahrscheinlichkeit für alle Werte. Wenn du aus früheren Tests weißt, dass „die Konversionsrate normalerweise zwischen 1,5 und 2,5 % liegt\", definierst du einen informativen Prior (Beta(15, 985)).",{"type":31,"tag":32,"props":119,"children":120},{},[121,126],{"type":31,"tag":54,"props":122,"children":123},{},[124],{"type":36,"value":125},"Likelihood:",{"type":36,"value":127}," Deine beobachteten Daten – z. B. 1000 Impressionen, 25 Konversionen.",{"type":31,"tag":32,"props":129,"children":130},{},[131,136,138,145],{"type":31,"tag":54,"props":132,"children":133},{},[134],{"type":36,"value":135},"Posterior:",{"type":36,"value":137}," Die aktualisierte Verteilung via Bayes-Theorem. Mit Beta-Binomial-Konjugation löst sich der Posterior analytisch: ",{"type":31,"tag":139,"props":140,"children":142},"code",{"className":141},[],[143],{"type":36,"value":144},"Beta(alpha + conversions, beta + non_conversions)",{"type":36,"value":146},".",{"type":31,"tag":32,"props":148,"children":149},{},[150,155],{"type":31,"tag":54,"props":151,"children":152},{},[153],{"type":36,"value":154},"Entscheidungsregel:",{"type":36,"value":156}," Du samplist die Posterior-Verteilungen von Variante A und B per Monte-Carlo-Simulation (z. B. 100.000 Iterationen). In jeder Iteration zählst du, wie oft B größer als A ist. Dieses Verhältnis ist „Wahrscheinlichkeit, dass B gewinnt\" (P(B > A)). Übersteigt diese Wahrscheinlichkeit 95 %, beendest du den Test und wählst B.",{"type":31,"tag":32,"props":158,"children":159},{},[160,165],{"type":31,"tag":54,"props":161,"children":162},{},[163],{"type":36,"value":164},"Sequentielle Überwachung:",{"type":36,"value":166}," Das Bayesian-Framework erlaubt dir, den Posterior täglich neu zu berechnen. Das „Peeking\"-Problem des Frequentisten gibt es nicht – Posterior-Updates sind ein natürlicher Teil der Bayesian-Inferenz. Jeden Morgen öffnest du das Dashboard und siehst aktuelle P(B > A) Werte: 65 % → 78 % → 89 % → 94 % → 96 %. Überschreitest du die 95 %-Schwelle, beendest du den Test.",{"type":31,"tag":32,"props":168,"children":169},{},[170],{"type":36,"value":171},"In der Praxis: Baseline 2 % Konversionsrate, Ziel 10 % relativer Lift (d. h. 2,2 %), 95 % Konfidenz-Schwelle. Der frequentistische Test braucht 15.000 Samples pro Variante (insgesamt 21 Tage). Der Bayesian-Test erreicht dieselbe Schwelle in 9–12 Tagen – weil das Prior-Wissen die Posterior schneller scharf macht.",{"type":31,"tag":173,"props":174,"children":176},"h3",{"id":175},"beispiel-simulationscode-python",[177],{"type":36,"value":178},"Beispiel-Simulationscode (Python)",{"type":31,"tag":180,"props":181,"children":185},"pre",{"className":182,"code":183,"language":184,"meta":15,"style":15},"language-python shiki shiki-themes github-dark","import numpy as np\nfrom scipy.stats import beta\n\n# Prior: Beta(1, 1) — uniform\nalpha_a, beta_a = 1, 1\nalpha_b, beta_b = 1, 1\n\n# Beobachtete Daten (Tag 5)\nviews_a, conv_a = 5000, 95\nviews_b, conv_b = 5000, 112\n\n# Posterior\npost_a = beta(alpha_a + conv_a, beta_a + views_a - conv_a)\npost_b = beta(alpha_b + conv_b, beta_b + views_b - conv_b)\n\n# Monte Carlo: P(B > A)\nsamples_a = post_a.rvs(100000)\nsamples_b = post_b.rvs(100000)\nprob_b_wins = (samples_b > samples_a).mean()\n\nprint(f\"P(B > A) = {prob_b_wins:.3f}\")\n# Ausgabe z. B.: P(B > A) = 0.923 → noch unter 95 %, Test fortsetzen\n","python",[186],{"type":31,"tag":139,"props":187,"children":188},{"__ignoreMap":15},[189,217,240,250,260,290,315,323,332,358,384,392,401,448,493,501,510,538,564,592,600,654],{"type":31,"tag":190,"props":191,"children":194},"span",{"class":192,"line":193},"line",1,[195,201,207,212],{"type":31,"tag":190,"props":196,"children":198},{"style":197},"--shiki-default:#F97583",[199],{"type":36,"value":200},"import",{"type":31,"tag":190,"props":202,"children":204},{"style":203},"--shiki-default:#E1E4E8",[205],{"type":36,"value":206}," numpy ",{"type":31,"tag":190,"props":208,"children":209},{"style":197},[210],{"type":36,"value":211},"as",{"type":31,"tag":190,"props":213,"children":214},{"style":203},[215],{"type":36,"value":216}," np\n",{"type":31,"tag":190,"props":218,"children":220},{"class":192,"line":219},2,[221,226,231,235],{"type":31,"tag":190,"props":222,"children":223},{"style":197},[224],{"type":36,"value":225},"from",{"type":31,"tag":190,"props":227,"children":228},{"style":203},[229],{"type":36,"value":230}," scipy.stats ",{"type":31,"tag":190,"props":232,"children":233},{"style":197},[234],{"type":36,"value":200},{"type":31,"tag":190,"props":236,"children":237},{"style":203},[238],{"type":36,"value":239}," beta\n",{"type":31,"tag":190,"props":241,"children":243},{"class":192,"line":242},3,[244],{"type":31,"tag":190,"props":245,"children":247},{"emptyLinePlaceholder":246},true,[248],{"type":36,"value":249},"\n",{"type":31,"tag":190,"props":251,"children":253},{"class":192,"line":252},4,[254],{"type":31,"tag":190,"props":255,"children":257},{"style":256},"--shiki-default:#6A737D",[258],{"type":36,"value":259},"# Prior: Beta(1, 1) — uniform\n",{"type":31,"tag":190,"props":261,"children":263},{"class":192,"line":262},5,[264,269,274,280,285],{"type":31,"tag":190,"props":265,"children":266},{"style":203},[267],{"type":36,"value":268},"alpha_a, beta_a ",{"type":31,"tag":190,"props":270,"children":271},{"style":197},[272],{"type":36,"value":273},"=",{"type":31,"tag":190,"props":275,"children":277},{"style":276},"--shiki-default:#79B8FF",[278],{"type":36,"value":279}," 1",{"type":31,"tag":190,"props":281,"children":282},{"style":203},[283],{"type":36,"value":284},", ",{"type":31,"tag":190,"props":286,"children":287},{"style":276},[288],{"type":36,"value":289},"1\n",{"type":31,"tag":190,"props":291,"children":293},{"class":192,"line":292},6,[294,299,303,307,311],{"type":31,"tag":190,"props":295,"children":296},{"style":203},[297],{"type":36,"value":298},"alpha_b, beta_b ",{"type":31,"tag":190,"props":300,"children":301},{"style":197},[302],{"type":36,"value":273},{"type":31,"tag":190,"props":304,"children":305},{"style":276},[306],{"type":36,"value":279},{"type":31,"tag":190,"props":308,"children":309},{"style":203},[310],{"type":36,"value":284},{"type":31,"tag":190,"props":312,"children":313},{"style":276},[314],{"type":36,"value":289},{"type":31,"tag":190,"props":316,"children":318},{"class":192,"line":317},7,[319],{"type":31,"tag":190,"props":320,"children":321},{"emptyLinePlaceholder":246},[322],{"type":36,"value":249},{"type":31,"tag":190,"props":324,"children":326},{"class":192,"line":325},8,[327],{"type":31,"tag":190,"props":328,"children":329},{"style":256},[330],{"type":36,"value":331},"# Beobachtete Daten (Tag 5)\n",{"type":31,"tag":190,"props":333,"children":334},{"class":192,"line":25},[335,340,344,349,353],{"type":31,"tag":190,"props":336,"children":337},{"style":203},[338],{"type":36,"value":339},"views_a, conv_a ",{"type":31,"tag":190,"props":341,"children":342},{"style":197},[343],{"type":36,"value":273},{"type":31,"tag":190,"props":345,"children":346},{"style":276},[347],{"type":36,"value":348}," 5000",{"type":31,"tag":190,"props":350,"children":351},{"style":203},[352],{"type":36,"value":284},{"type":31,"tag":190,"props":354,"children":355},{"style":276},[356],{"type":36,"value":357},"95\n",{"type":31,"tag":190,"props":359,"children":361},{"class":192,"line":360},10,[362,367,371,375,379],{"type":31,"tag":190,"props":363,"children":364},{"style":203},[365],{"type":36,"value":366},"views_b, conv_b ",{"type":31,"tag":190,"props":368,"children":369},{"style":197},[370],{"type":36,"value":273},{"type":31,"tag":190,"props":372,"children":373},{"style":276},[374],{"type":36,"value":348},{"type":31,"tag":190,"props":376,"children":377},{"style":203},[378],{"type":36,"value":284},{"type":31,"tag":190,"props":380,"children":381},{"style":276},[382],{"type":36,"value":383},"112\n",{"type":31,"tag":190,"props":385,"children":387},{"class":192,"line":386},11,[388],{"type":31,"tag":190,"props":389,"children":390},{"emptyLinePlaceholder":246},[391],{"type":36,"value":249},{"type":31,"tag":190,"props":393,"children":395},{"class":192,"line":394},12,[396],{"type":31,"tag":190,"props":397,"children":398},{"style":256},[399],{"type":36,"value":400},"# Posterior\n",{"type":31,"tag":190,"props":402,"children":404},{"class":192,"line":403},13,[405,410,414,419,424,429,433,438,443],{"type":31,"tag":190,"props":406,"children":407},{"style":203},[408],{"type":36,"value":409},"post_a ",{"type":31,"tag":190,"props":411,"children":412},{"style":197},[413],{"type":36,"value":273},{"type":31,"tag":190,"props":415,"children":416},{"style":203},[417],{"type":36,"value":418}," beta(alpha_a ",{"type":31,"tag":190,"props":420,"children":421},{"style":197},[422],{"type":36,"value":423},"+",{"type":31,"tag":190,"props":425,"children":426},{"style":203},[427],{"type":36,"value":428}," conv_a, beta_a ",{"type":31,"tag":190,"props":430,"children":431},{"style":197},[432],{"type":36,"value":423},{"type":31,"tag":190,"props":434,"children":435},{"style":203},[436],{"type":36,"value":437}," views_a ",{"type":31,"tag":190,"props":439,"children":440},{"style":197},[441],{"type":36,"value":442},"-",{"type":31,"tag":190,"props":444,"children":445},{"style":203},[446],{"type":36,"value":447}," conv_a)\n",{"type":31,"tag":190,"props":449,"children":451},{"class":192,"line":450},14,[452,457,461,466,470,475,479,484,488],{"type":31,"tag":190,"props":453,"children":454},{"style":203},[455],{"type":36,"value":456},"post_b ",{"type":31,"tag":190,"props":458,"children":459},{"style":197},[460],{"type":36,"value":273},{"type":31,"tag":190,"props":462,"children":463},{"style":203},[464],{"type":36,"value":465}," beta(alpha_b ",{"type":31,"tag":190,"props":467,"children":468},{"style":197},[469],{"type":36,"value":423},{"type":31,"tag":190,"props":471,"children":472},{"style":203},[473],{"type":36,"value":474}," conv_b, beta_b ",{"type":31,"tag":190,"props":476,"children":477},{"style":197},[478],{"type":36,"value":423},{"type":31,"tag":190,"props":480,"children":481},{"style":203},[482],{"type":36,"value":483}," views_b ",{"type":31,"tag":190,"props":485,"children":486},{"style":197},[487],{"type":36,"value":442},{"type":31,"tag":190,"props":489,"children":490},{"style":203},[491],{"type":36,"value":492}," conv_b)\n",{"type":31,"tag":190,"props":494,"children":496},{"class":192,"line":495},15,[497],{"type":31,"tag":190,"props":498,"children":499},{"emptyLinePlaceholder":246},[500],{"type":36,"value":249},{"type":31,"tag":190,"props":502,"children":504},{"class":192,"line":503},16,[505],{"type":31,"tag":190,"props":506,"children":507},{"style":256},[508],{"type":36,"value":509},"# Monte Carlo: P(B > A)\n",{"type":31,"tag":190,"props":511,"children":513},{"class":192,"line":512},17,[514,519,523,528,533],{"type":31,"tag":190,"props":515,"children":516},{"style":203},[517],{"type":36,"value":518},"samples_a ",{"type":31,"tag":190,"props":520,"children":521},{"style":197},[522],{"type":36,"value":273},{"type":31,"tag":190,"props":524,"children":525},{"style":203},[526],{"type":36,"value":527}," post_a.rvs(",{"type":31,"tag":190,"props":529,"children":530},{"style":276},[531],{"type":36,"value":532},"100000",{"type":31,"tag":190,"props":534,"children":535},{"style":203},[536],{"type":36,"value":537},")\n",{"type":31,"tag":190,"props":539,"children":541},{"class":192,"line":540},18,[542,547,551,556,560],{"type":31,"tag":190,"props":543,"children":544},{"style":203},[545],{"type":36,"value":546},"samples_b ",{"type":31,"tag":190,"props":548,"children":549},{"style":197},[550],{"type":36,"value":273},{"type":31,"tag":190,"props":552,"children":553},{"style":203},[554],{"type":36,"value":555}," post_b.rvs(",{"type":31,"tag":190,"props":557,"children":558},{"style":276},[559],{"type":36,"value":532},{"type":31,"tag":190,"props":561,"children":562},{"style":203},[563],{"type":36,"value":537},{"type":31,"tag":190,"props":565,"children":567},{"class":192,"line":566},19,[568,573,577,582,587],{"type":31,"tag":190,"props":569,"children":570},{"style":203},[571],{"type":36,"value":572},"prob_b_wins ",{"type":31,"tag":190,"props":574,"children":575},{"style":197},[576],{"type":36,"value":273},{"type":31,"tag":190,"props":578,"children":579},{"style":203},[580],{"type":36,"value":581}," (samples_b ",{"type":31,"tag":190,"props":583,"children":584},{"style":197},[585],{"type":36,"value":586},">",{"type":31,"tag":190,"props":588,"children":589},{"style":203},[590],{"type":36,"value":591}," samples_a).mean()\n",{"type":31,"tag":190,"props":593,"children":595},{"class":192,"line":594},20,[596],{"type":31,"tag":190,"props":597,"children":598},{"emptyLinePlaceholder":246},[599],{"type":36,"value":249},{"type":31,"tag":190,"props":601,"children":603},{"class":192,"line":602},21,[604,609,614,619,625,630,635,640,645,650],{"type":31,"tag":190,"props":605,"children":606},{"style":276},[607],{"type":36,"value":608},"print",{"type":31,"tag":190,"props":610,"children":611},{"style":203},[612],{"type":36,"value":613},"(",{"type":31,"tag":190,"props":615,"children":616},{"style":197},[617],{"type":36,"value":618},"f",{"type":31,"tag":190,"props":620,"children":622},{"style":621},"--shiki-default:#9ECBFF",[623],{"type":36,"value":624},"\"P(B > A) = ",{"type":31,"tag":190,"props":626,"children":627},{"style":276},[628],{"type":36,"value":629},"{",{"type":31,"tag":190,"props":631,"children":632},{"style":203},[633],{"type":36,"value":634},"prob_b_wins",{"type":31,"tag":190,"props":636,"children":637},{"style":197},[638],{"type":36,"value":639},":.3f",{"type":31,"tag":190,"props":641,"children":642},{"style":276},[643],{"type":36,"value":644},"}",{"type":31,"tag":190,"props":646,"children":647},{"style":621},[648],{"type":36,"value":649},"\"",{"type":31,"tag":190,"props":651,"children":652},{"style":203},[653],{"type":36,"value":537},{"type":31,"tag":190,"props":655,"children":657},{"class":192,"line":656},22,[658],{"type":31,"tag":190,"props":659,"children":660},{"style":256},[661],{"type":36,"value":662},"# Ausgabe z. B.: P(B > A) = 0.923 → noch unter 95 %, Test fortsetzen\n",{"type":31,"tag":39,"props":664,"children":666},{"id":665},"sample-size-dynamik-und-kriterien-für-vorzeitige-beendigung",[667],{"type":36,"value":668},"Sample-Size-Dynamik und Kriterien für vorzeitige Beendigung",{"type":31,"tag":32,"props":670,"children":671},{},[672],{"type":36,"value":673},"Der Geschwindigkeitsvorteil des Bayesian-Tests kommt aus der dynamischen Stichprobengröße. Statt eines festen N-Ziels bindest du die Stopping-Rule an Posterior-Konfidenz. Zwei gängige Kriterien:",{"type":31,"tag":32,"props":675,"children":676},{},[677,682],{"type":31,"tag":54,"props":678,"children":679},{},[680],{"type":36,"value":681},"Wahrscheinlichkeitsschwelle:",{"type":36,"value":683}," P(B > A) ≥ 0,95, dann stoppen. Das bedeutet: „Die Wahrscheinlichkeit, dass B wirklich besser ist, liegt bei 95 %.\" Einige Teams nutzen 99 % (konservativer), andere 90 % (aggressiver – für Test-Velocity).",{"type":31,"tag":32,"props":685,"children":686},{},[687,692,694,699],{"type":31,"tag":54,"props":688,"children":689},{},[690],{"type":36,"value":691},"Erwarteter Verlust:",{"type":36,"value":693}," Wenn du B wählst und A war in Wirklichkeit besser, wie groß ist dein Verlust? Expected Loss = E",{"type":31,"tag":190,"props":695,"children":696},{},[697],{"type":36,"value":698},"max(0, A - B)",{"type":36,"value":700},". Liegt dieser Verlust unter akzeptablem Niveau (z. B. \u003C 0,0001 absolute Konversionsrate-Differenz), beendest du den Test. Diese Metrik verwaltet Risiko aus der Perspektive „Kosten einer falschen Entscheidung.\"",{"type":31,"tag":32,"props":702,"children":703},{},[704,709],{"type":31,"tag":54,"props":705,"children":706},{},[707],{"type":36,"value":708},"Minimale Sample-Grenze:",{"type":36,"value":710}," Um vorzeitiges Stoppen zu bremsen: „Sampel mindestens 3000, dann Bayesian Stopping-Rule anwenden.\" Das verhindert, dass der Prior zu dominant wird.",{"type":31,"tag":32,"props":712,"children":713},{},[714],{"type":36,"value":715},"Beispiel-Szenario: E-Commerce-Checkout-CTA-Farb-Test (grün vs. orange). Baseline 3,2 % Konversion. Woche 1: 8000 Views, P(orange > grün) = 87 %. Woche 2: 16.000 Views, P = 94 %. Woche 3, Tag 2 (insgesamt 18.500 Views), P = 96 %. Der frequentistische Test hätte 25.000 Views verlangt (insgesamt 18 Tage) – du hast nach 10 Tagen gestoppt. Du hast die Testdauer um 44 % gekürzt.",{"type":31,"tag":32,"props":717,"children":718},{},[719],{"type":36,"value":720},"Trade-off: Vorzeitiges Stoppen kann das Risiko erhöhen, „zufällig gut startende, aber später regedierende\" Varianten zu wählen. Zur Risikoreduktion: (1) Lege eine minimale Sample-Grenze fest, (2) Bei kleinen Effektgrößen (z. B. 5 % relativer Lift) erhöhe den Threshold auf 99 %, (3) Beobachte die Standard-Abweichung des Posterior – ist sie noch groß (hohe Unsicherheit), sampel mehr.",{"type":31,"tag":39,"props":722,"children":724},{"id":723},"prior-wahl-und-wissensspeicherung",[725],{"type":36,"value":726},"Prior-Wahl und Wissensspeicherung",{"type":31,"tag":32,"props":728,"children":729},{},[730],{"type":36,"value":731},"Die Kraft des Bayesian-Tests kommt aus der Formalisierung von Prior-Wissen. Aber falsche Prior-Wahl erzeugt Bias. Zwei Extreme:",{"type":31,"tag":32,"props":733,"children":734},{},[735,740],{"type":31,"tag":54,"props":736,"children":737},{},[738],{"type":36,"value":739},"Non-informativer Prior (Beta(1,1)):",{"type":36,"value":741}," Keine Vorwissens-Annahme. Jeder Test startet auf einer leeren Tafel. Vorteil: unvoreingenommen. Nachteil: Es braucht mehr Daten, um den Posterior sharp zu machen – ähnlich frequentistischer Sample-Size.",{"type":31,"tag":32,"props":743,"children":744},{},[745,750],{"type":31,"tag":54,"props":746,"children":747},{},[748],{"type":36,"value":749},"Informativer Prior (Beta(α, β)):",{"type":36,"value":751}," Du transportierst Wissen aus früheren Tests, Sektorbenchmarks oder Baseline. Beispiel: „CTA-Button-Tests zeigen normalerweise 2–4 % Konversionsrate, Mittelwert 2,8 %\" → defin Beta(28, 972) Prior (Mittelwert 2,8 %, Varianz angemessen).",{"type":31,"tag":32,"props":753,"children":754},{},[755],{"type":36,"value":756},"Informativer Prior verkürzt die Testdauer, weil Prior + neue Daten schnellere Konvergenz liefern. Aber Risiko: Ist der Prior falsch (z. B. kopiert aus einem alten Vertical, neuer Segment ist anders), wird der Posterior biased. Zwei Schutzmaßnahmen:",{"type":31,"tag":32,"props":758,"children":759},{},[760,765],{"type":31,"tag":54,"props":761,"children":762},{},[763],{"type":36,"value":764},"Prior-Sensitivitätsanalyse:",{"type":36,"value":766}," Führe den Test mit verschiedenen Prioren aus (schwach, mittel, stark informativ) und prüfe, ob die Ergebnisse sich stark verändern. Wenn mit schwachem Prior 60 % Gewinn-Wahrscheinlichkeit und mit starkem 98 %, dann ist der Test zu prior-abhängig – verlängere ihn. Die Daten können den Prior noch nicht überrollen.",{"type":31,"tag":32,"props":768,"children":769},{},[770,775],{"type":31,"tag":54,"props":771,"children":772},{},[773],{"type":36,"value":774},"Hierarchisches Prior-Modell:",{"type":36,"value":776}," Bei Multiple-Segment-Tests (mobil vs. Desktop, Land-basiert) verwende hierarchisches Bayesian-Modell. Jedes Segment hat seine eigene Konversionsrate, aber ein globaler Prior schrumpft es zum Population-Mittelwert. Das reduziert Segment-Level Overfitting.",{"type":31,"tag":32,"props":778,"children":779},{},[780],{"type":36,"value":781},"Praktischer Rat: Führe die ersten 5–10 Tests mit non-informativem Prior aus, sammle Ergebnisse, berechne Mittelwert und Varianz, nutze das in nachfolgenden Tests als informativen Prior. Dieses „Meta-Learning\"-Vorgehen speichert kumulatives Test-Wissen.",{"type":31,"tag":39,"props":783,"children":785},{"id":784},"organisatorische-integration-und-entscheidungsprotokoll",[786],{"type":36,"value":787},"Organisatorische Integration und Entscheidungsprotokoll",{"type":31,"tag":32,"props":789,"children":790},{},[791],{"type":36,"value":792},"Bayesian A\u002FB-Test in die Teamkultur zu integrieren ist organisatorisch, nicht technisch. Wenn du einem Frequentist-gewöhnten Team sagst „jetzt könnt ihr täglich schauen\", ist die erste Reaktion gemischt: „Wo ist der p-Wert?\" Zwei Schritte:",{"type":31,"tag":32,"props":794,"children":795},{},[796,801],{"type":31,"tag":54,"props":797,"children":798},{},[799],{"type":36,"value":800},"Training + Onboarding:",{"type":36,"value":802}," Erkläre, was P(B > A) bedeutet. Du kannst sagen: „95 % Wahrscheinlichkeit, dass B besser ist\" – klare Sprache statt frequentistischer Indirektion „p \u003C 0,05 also Null abgelehnt.\" Führe die ersten 2–3 Tests parallel durch – Frequentist und Bayesian analysieren, vergleich zeigen. Das Team sieht den Unterschied und adoptiert schneller.",{"type":31,"tag":32,"props":804,"children":805},{},[806,811],{"type":31,"tag":54,"props":807,"children":808},{},[809],{"type":36,"value":810},"Decision-Threshold-Standardisierung:",{"type":36,"value":812}," Bei welcher Wahrscheinlichkeit beendest du den Test? 95 %, 99 %? Das hängt von Risk-Toleranz ab. High-Traffic + niedriges Risiko (z. B. Email Subject Line) → 90 % reicht. Low-Traffic + hohes Risiko (z. B. Pricing-Page Redesign) → 99 % nutzen. Schreib diese Thresholds in dein Test Playbook.",{"type":31,"tag":32,"props":814,"children":815},{},[816,821],{"type":31,"tag":54,"props":817,"children":818},{},[819],{"type":36,"value":820},"Post-Test-Monitoring:",{"type":36,"value":822}," Du hast den Test beendet, B als Gewinner erklärt. Aber 2 Wochen nach vollständigem Rollout sinkt die Konversionsrate – Regression zur Mitte oder externer Faktor (Kampagne, Saisonalität). Bayesian-Test mindert dieses Risiko, hebt es nicht auf. Lösung: 1 Woche Post-Rollout Monitoring, wenn der Posterior-Mittelwert > 10 % sinkt, Rollback triggern.",{"type":31,"tag":32,"props":824,"children":825},{},[826,831],{"type":31,"tag":54,"props":827,"children":828},{},[829],{"type":36,"value":830},"Tooling:",{"type":36,"value":832}," Google Optimize bietet Bayesian-Modus, aber begrenzt. VWO, Optimizely haben teilweise Support. Custom Stack: Python (PyMC3, ArviZ) + BigQuery + Looker-Dashboard. Täglich updatet ein Airflow-Job die Posterior-Werte, Looker zeigt P(B > A)-Metrik. Slack-Alert, wenn Threshold überschritten.",{"type":31,"tag":834,"props":835,"children":836},"hr",{},[],{"type":31,"tag":32,"props":838,"children":839},{},[840],{"type":36,"value":841},"Der Bayesian A\u002FB-Test erhöht Test-Velocity, benötigt aber stat",{"type":31,"tag":843,"props":844,"children":845},"style",{},[846],{"type":36,"value":847},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":15,"searchDepth":242,"depth":242,"links":849},[850,851,854,855,856],{"id":41,"depth":219,"text":44},{"id":99,"depth":219,"text":102,"children":852},[853],{"id":175,"depth":242,"text":178},{"id":665,"depth":219,"text":668},{"id":723,"depth":219,"text":726},{"id":784,"depth":219,"text":787},"markdown","content:de:marketing:bayesian-ab-test-schnelle-entscheidungen.md","content","de\u002Fmarketing\u002Fbayesian-ab-test-schnelle-entscheidungen.md","de\u002Fmarketing\u002Fbayesian-ab-test-schnelle-entscheidungen","md",1781791478236]