[{"data":1,"prerenderedAt":890},["ShallowReactive",2],{"article-alternates":3,"article-\u002Fes\u002Fmarketing\u002Fprueba-bayesiana-ab-toma-rapida-decisiones":13},{"i18nKey":4,"paths":5},"marketing-002-2026-06",{"de":6,"en":7,"es":8,"fr":9,"it":10,"ru":11,"tr":12},"\u002Fde\u002Fmarketing\u002Fbayesian-ab-test-schnelle-entscheidungsfindung","\u002Fen\u002Fmarketing\u002Ffast-decision-making-with-bayesian-ab-testing","\u002Fes\u002Fmarketing\u002Fprueba-bayesiana-ab-toma-rapida-decisiones","\u002Ffr\u002Fmarketing\u002Fbayesian-ab-testi-hizli-karar-verme","\u002Fit\u002Fmarketing\u002Ftest-bayesiano-ab-decisione-veloce","\u002Fru\u002Fmarketing\u002Fbayes-a-b-testi-hizli-karar-verme","\u002Ftr\u002Fmarketing\u002Fbayesian-a-b-test-ile-hizli-karar-verme",{"_path":8,"_dir":14,"_draft":15,"_partial":15,"_locale":16,"title":17,"description":18,"publishedAt":19,"modifiedAt":19,"category":14,"i18nKey":4,"tags":20,"readingTime":26,"author":27,"body":28,"_type":884,"_id":885,"_source":886,"_file":887,"_stem":888,"_extension":889},"marketing",false,"","Test A\u002FB Bayesiano: Toma de Decisiones Rápida","Reemplaza las reglas rígidas de tamaño muestral frequentista con enfoque Bayesiano para pruebas secuenciales. Actualiza distribuciones de probabilidad en tiempo real y detén antes.","2026-06-18",[21,22,23,24,25],"ab-testing","estadistica-bayesiana","optimizacion-conversion","prueba-secuencial","performance-marketing",8,"Roibase",{"type":29,"children":30,"toc":873},"root",[31,39,46,51,56,61,67,87,101,106,111,118,554,567,573,578,583,618,634,640,645,650,757,762,768,778,788,798,808,814,819,824,848,853,858,862,867],{"type":32,"tag":33,"props":34,"children":35},"element","p",{},[36],{"type":37,"value":38},"text","El test A\u002FB clásico depende de un tamaño muestral fijo. Esperas a alcanzar N usuarios, ejecutas un t-test, controlas el p-value. Pero la realidad del mercado es implacable: si la variante B pierde claramente cada día, quemar tráfico durante 2 semanas más es desperdicio. El enfoque Bayesiano resuelve esto — actualiza la distribución posterior cada día y afirmas \"la probabilidad de que la variante A gane es 94%\". Defines el umbral de decisión, no estás atrapado en la rigidez frequentista de p\u003C0.05.",{"type":32,"tag":40,"props":41,"children":43},"h2",{"id":42},"las-limitaciones-estructurales-del-test-frequentista",[44],{"type":37,"value":45},"Las Limitaciones Estructurales del Test Frequentista",{"type":32,"tag":33,"props":47,"children":48},{},[49],{"type":37,"value":50},"El test A\u002FB tradicional se basa en el marco Neyman-Pearson. Defines la hipótesis nula (H₀: sin diferencia entre variantes), estableces el nivel alpha (típicamente 0.05), determinas el efecto detectable mínimo (MDE), realizas análisis de potencia (80%), y esperas hasta alcanzar el tamaño muestral resultante. Hacer un peek antes de terminar el test infla el error Tipo I — por eso el \"peeking\" está prohibido.",{"type":32,"tag":33,"props":52,"children":53},{},[54],{"type":37,"value":55},"El problema: en campañas digitales, el tráfico cuesta dinero cada día. Si el cálculo de tamaño muestral dice 12.000 usuarios y recibes 800 diarios, esperas 15 días. Pero en el día 5, la tasa de conversión de la variante B cae de 2.1% a 1.3% y aún quemas 10 días más. La metodología frequentista lo justifica porque \"detención temprana = sesgo\". En realidad, el escenario de prueba no es estático — presupuesto limitado, estacionalidad, competencia que se mueve. Las reglas rígidas de tamaño muestral no dejan espacio para flexibilidad.",{"type":32,"tag":33,"props":57,"children":58},{},[59],{"type":37,"value":60},"Hay otro problema: el p-value solo dice \"si H₀ fuera cierta, ¿cuál es la probabilidad de ver estos datos?\". No te dice la probabilidad de que la variante A sea realmente mejor. Si p=0.03, rechazas H₀, pero no puedes afirmar \"A tiene 97% de probabilidad de vencer a B\". El lenguaje frequentista solo te da \"significancia estadística\" — insuficiente para decidir en negocio.",{"type":32,"tag":40,"props":62,"children":64},{"id":63},"la-lógica-del-enfoque-bayesiano",[65],{"type":37,"value":66},"La Lógica del Enfoque Bayesiano",{"type":32,"tag":33,"props":68,"children":69},{},[70,72,78,80,85],{"type":37,"value":71},"El marco Bayesiano convierte información anterior en distribución posterior. ",{"type":32,"tag":73,"props":74,"children":75},"strong",{},[76],{"type":37,"value":77},"Prior",{"type":37,"value":79},": tu creencia sobre la tasa de conversión antes de la prueba. Conforme llegan datos, el teorema de Bayes actualiza el prior. ",{"type":32,"tag":73,"props":81,"children":82},{},[83],{"type":37,"value":84},"Posterior",{"type":37,"value":86},": la distribución probable de la tasa de conversión según los datos acumulados.",{"type":32,"tag":33,"props":88,"children":89},{},[90,92,96],{"type":37,"value":91},"Fórmula:",{"type":32,"tag":93,"props":94,"children":95},"br",{},[],{"type":32,"tag":73,"props":97,"children":98},{},[99],{"type":37,"value":100},"P(θ | data) ∝ P(data | θ) × P(θ)",{"type":32,"tag":33,"props":102,"children":103},{},[104],{"type":37,"value":105},"θ = tasa de conversión, data = conversiones y no-conversiones observadas. Likelihood (probabilidad de datos) × prior → posterior. La distribución Beta es el prior conjugado, así que el cálculo es simple: si la variante A muestra α conversiones y β no-conversiones, posterior = Beta(α+1, β+1).",{"type":32,"tag":33,"props":107,"children":108},{},[109],{"type":37,"value":110},"Cada día actualizas el posterior con datos nuevos. La ventaja crítica de la prueba secuencial es esta: comparas las distribuciones posteriores y calculas \"la probabilidad de que la tasa de conversión de A sea superior a la de B\" mediante simulación Monte Carlo. Si supera 95%, decides. No es \"alcanza N, luego mira\" como en frequentista, sino \"mira cada día, decide cuando cruzas el umbral\".",{"type":32,"tag":112,"props":113,"children":115},"h3",{"id":114},"ejemplo-de-cálculo-posterior",[116],{"type":37,"value":117},"Ejemplo de Cálculo Posterior",{"type":32,"tag":119,"props":120,"children":124},"pre",{"className":121,"code":122,"language":123,"meta":16,"style":16},"language-python shiki shiki-themes github-dark","import numpy as np\nfrom scipy.stats import beta\n\n# Variante A: 120 conversiones, 1200 impresiones\nalpha_A = 120 + 1  # +1 para prior uniforme\nbeta_A = (1200 - 120) + 1\n\n# Variante B: 95 conversiones, 1150 impresiones\nalpha_B = 95 + 1\nbeta_B = (1150 - 95) + 1\n\n# Monte Carlo: extrae 10,000 muestras\nsamples_A = beta.rvs(alpha_A, beta_A, size=10000)\nsamples_B = beta.rvs(alpha_B, beta_B, size=10000)\n\n# Probabilidad de que A > B\nprob_A_wins = (samples_A > samples_B).mean()\nprint(f\"P(A > B) = {prob_A_wins:.3f}\")\n","python",[125],{"type":32,"tag":126,"props":127,"children":128},"code",{"__ignoreMap":16},[129,157,180,190,200,235,282,290,298,324,366,374,383,421,455,463,472,500],{"type":32,"tag":130,"props":131,"children":134},"span",{"class":132,"line":133},"line",1,[135,141,147,152],{"type":32,"tag":130,"props":136,"children":138},{"style":137},"--shiki-default:#F97583",[139],{"type":37,"value":140},"import",{"type":32,"tag":130,"props":142,"children":144},{"style":143},"--shiki-default:#E1E4E8",[145],{"type":37,"value":146}," numpy ",{"type":32,"tag":130,"props":148,"children":149},{"style":137},[150],{"type":37,"value":151},"as",{"type":32,"tag":130,"props":153,"children":154},{"style":143},[155],{"type":37,"value":156}," np\n",{"type":32,"tag":130,"props":158,"children":160},{"class":132,"line":159},2,[161,166,171,175],{"type":32,"tag":130,"props":162,"children":163},{"style":137},[164],{"type":37,"value":165},"from",{"type":32,"tag":130,"props":167,"children":168},{"style":143},[169],{"type":37,"value":170}," scipy.stats ",{"type":32,"tag":130,"props":172,"children":173},{"style":137},[174],{"type":37,"value":140},{"type":32,"tag":130,"props":176,"children":177},{"style":143},[178],{"type":37,"value":179}," beta\n",{"type":32,"tag":130,"props":181,"children":183},{"class":132,"line":182},3,[184],{"type":32,"tag":130,"props":185,"children":187},{"emptyLinePlaceholder":186},true,[188],{"type":37,"value":189},"\n",{"type":32,"tag":130,"props":191,"children":193},{"class":132,"line":192},4,[194],{"type":32,"tag":130,"props":195,"children":197},{"style":196},"--shiki-default:#6A737D",[198],{"type":37,"value":199},"# Variante A: 120 conversiones, 1200 impresiones\n",{"type":32,"tag":130,"props":201,"children":203},{"class":132,"line":202},5,[204,209,214,220,225,230],{"type":32,"tag":130,"props":205,"children":206},{"style":143},[207],{"type":37,"value":208},"alpha_A ",{"type":32,"tag":130,"props":210,"children":211},{"style":137},[212],{"type":37,"value":213},"=",{"type":32,"tag":130,"props":215,"children":217},{"style":216},"--shiki-default:#79B8FF",[218],{"type":37,"value":219}," 120",{"type":32,"tag":130,"props":221,"children":222},{"style":137},[223],{"type":37,"value":224}," +",{"type":32,"tag":130,"props":226,"children":227},{"style":216},[228],{"type":37,"value":229}," 1",{"type":32,"tag":130,"props":231,"children":232},{"style":196},[233],{"type":37,"value":234},"  # +1 para prior uniforme\n",{"type":32,"tag":130,"props":236,"children":238},{"class":132,"line":237},6,[239,244,248,253,258,263,267,272,277],{"type":32,"tag":130,"props":240,"children":241},{"style":143},[242],{"type":37,"value":243},"beta_A ",{"type":32,"tag":130,"props":245,"children":246},{"style":137},[247],{"type":37,"value":213},{"type":32,"tag":130,"props":249,"children":250},{"style":143},[251],{"type":37,"value":252}," (",{"type":32,"tag":130,"props":254,"children":255},{"style":216},[256],{"type":37,"value":257},"1200",{"type":32,"tag":130,"props":259,"children":260},{"style":137},[261],{"type":37,"value":262}," -",{"type":32,"tag":130,"props":264,"children":265},{"style":216},[266],{"type":37,"value":219},{"type":32,"tag":130,"props":268,"children":269},{"style":143},[270],{"type":37,"value":271},") ",{"type":32,"tag":130,"props":273,"children":274},{"style":137},[275],{"type":37,"value":276},"+",{"type":32,"tag":130,"props":278,"children":279},{"style":216},[280],{"type":37,"value":281}," 1\n",{"type":32,"tag":130,"props":283,"children":285},{"class":132,"line":284},7,[286],{"type":32,"tag":130,"props":287,"children":288},{"emptyLinePlaceholder":186},[289],{"type":37,"value":189},{"type":32,"tag":130,"props":291,"children":292},{"class":132,"line":26},[293],{"type":32,"tag":130,"props":294,"children":295},{"style":196},[296],{"type":37,"value":297},"# Variante B: 95 conversiones, 1150 impresiones\n",{"type":32,"tag":130,"props":299,"children":301},{"class":132,"line":300},9,[302,307,311,316,320],{"type":32,"tag":130,"props":303,"children":304},{"style":143},[305],{"type":37,"value":306},"alpha_B ",{"type":32,"tag":130,"props":308,"children":309},{"style":137},[310],{"type":37,"value":213},{"type":32,"tag":130,"props":312,"children":313},{"style":216},[314],{"type":37,"value":315}," 95",{"type":32,"tag":130,"props":317,"children":318},{"style":137},[319],{"type":37,"value":224},{"type":32,"tag":130,"props":321,"children":322},{"style":216},[323],{"type":37,"value":281},{"type":32,"tag":130,"props":325,"children":327},{"class":132,"line":326},10,[328,333,337,341,346,350,354,358,362],{"type":32,"tag":130,"props":329,"children":330},{"style":143},[331],{"type":37,"value":332},"beta_B ",{"type":32,"tag":130,"props":334,"children":335},{"style":137},[336],{"type":37,"value":213},{"type":32,"tag":130,"props":338,"children":339},{"style":143},[340],{"type":37,"value":252},{"type":32,"tag":130,"props":342,"children":343},{"style":216},[344],{"type":37,"value":345},"1150",{"type":32,"tag":130,"props":347,"children":348},{"style":137},[349],{"type":37,"value":262},{"type":32,"tag":130,"props":351,"children":352},{"style":216},[353],{"type":37,"value":315},{"type":32,"tag":130,"props":355,"children":356},{"style":143},[357],{"type":37,"value":271},{"type":32,"tag":130,"props":359,"children":360},{"style":137},[361],{"type":37,"value":276},{"type":32,"tag":130,"props":363,"children":364},{"style":216},[365],{"type":37,"value":281},{"type":32,"tag":130,"props":367,"children":369},{"class":132,"line":368},11,[370],{"type":32,"tag":130,"props":371,"children":372},{"emptyLinePlaceholder":186},[373],{"type":37,"value":189},{"type":32,"tag":130,"props":375,"children":377},{"class":132,"line":376},12,[378],{"type":32,"tag":130,"props":379,"children":380},{"style":196},[381],{"type":37,"value":382},"# Monte Carlo: extrae 10,000 muestras\n",{"type":32,"tag":130,"props":384,"children":386},{"class":132,"line":385},13,[387,392,396,401,407,411,416],{"type":32,"tag":130,"props":388,"children":389},{"style":143},[390],{"type":37,"value":391},"samples_A ",{"type":32,"tag":130,"props":393,"children":394},{"style":137},[395],{"type":37,"value":213},{"type":32,"tag":130,"props":397,"children":398},{"style":143},[399],{"type":37,"value":400}," beta.rvs(alpha_A, beta_A, ",{"type":32,"tag":130,"props":402,"children":404},{"style":403},"--shiki-default:#FFAB70",[405],{"type":37,"value":406},"size",{"type":32,"tag":130,"props":408,"children":409},{"style":137},[410],{"type":37,"value":213},{"type":32,"tag":130,"props":412,"children":413},{"style":216},[414],{"type":37,"value":415},"10000",{"type":32,"tag":130,"props":417,"children":418},{"style":143},[419],{"type":37,"value":420},")\n",{"type":32,"tag":130,"props":422,"children":424},{"class":132,"line":423},14,[425,430,434,439,443,447,451],{"type":32,"tag":130,"props":426,"children":427},{"style":143},[428],{"type":37,"value":429},"samples_B ",{"type":32,"tag":130,"props":431,"children":432},{"style":137},[433],{"type":37,"value":213},{"type":32,"tag":130,"props":435,"children":436},{"style":143},[437],{"type":37,"value":438}," beta.rvs(alpha_B, beta_B, ",{"type":32,"tag":130,"props":440,"children":441},{"style":403},[442],{"type":37,"value":406},{"type":32,"tag":130,"props":444,"children":445},{"style":137},[446],{"type":37,"value":213},{"type":32,"tag":130,"props":448,"children":449},{"style":216},[450],{"type":37,"value":415},{"type":32,"tag":130,"props":452,"children":453},{"style":143},[454],{"type":37,"value":420},{"type":32,"tag":130,"props":456,"children":458},{"class":132,"line":457},15,[459],{"type":32,"tag":130,"props":460,"children":461},{"emptyLinePlaceholder":186},[462],{"type":37,"value":189},{"type":32,"tag":130,"props":464,"children":466},{"class":132,"line":465},16,[467],{"type":32,"tag":130,"props":468,"children":469},{"style":196},[470],{"type":37,"value":471},"# Probabilidad de que A > B\n",{"type":32,"tag":130,"props":473,"children":475},{"class":132,"line":474},17,[476,481,485,490,495],{"type":32,"tag":130,"props":477,"children":478},{"style":143},[479],{"type":37,"value":480},"prob_A_wins ",{"type":32,"tag":130,"props":482,"children":483},{"style":137},[484],{"type":37,"value":213},{"type":32,"tag":130,"props":486,"children":487},{"style":143},[488],{"type":37,"value":489}," (samples_A ",{"type":32,"tag":130,"props":491,"children":492},{"style":137},[493],{"type":37,"value":494},">",{"type":32,"tag":130,"props":496,"children":497},{"style":143},[498],{"type":37,"value":499}," samples_B).mean()\n",{"type":32,"tag":130,"props":501,"children":503},{"class":132,"line":502},18,[504,509,514,519,525,530,535,540,545,550],{"type":32,"tag":130,"props":505,"children":506},{"style":216},[507],{"type":37,"value":508},"print",{"type":32,"tag":130,"props":510,"children":511},{"style":143},[512],{"type":37,"value":513},"(",{"type":32,"tag":130,"props":515,"children":516},{"style":137},[517],{"type":37,"value":518},"f",{"type":32,"tag":130,"props":520,"children":522},{"style":521},"--shiki-default:#9ECBFF",[523],{"type":37,"value":524},"\"P(A > B) = ",{"type":32,"tag":130,"props":526,"children":527},{"style":216},[528],{"type":37,"value":529},"{",{"type":32,"tag":130,"props":531,"children":532},{"style":143},[533],{"type":37,"value":534},"prob_A_wins",{"type":32,"tag":130,"props":536,"children":537},{"style":137},[538],{"type":37,"value":539},":.3f",{"type":32,"tag":130,"props":541,"children":542},{"style":216},[543],{"type":37,"value":544},"}",{"type":32,"tag":130,"props":546,"children":547},{"style":521},[548],{"type":37,"value":549},"\"",{"type":32,"tag":130,"props":551,"children":552},{"style":143},[553],{"type":37,"value":420},{"type":32,"tag":33,"props":555,"children":556},{},[557,559,565],{"type":37,"value":558},"Salida de ejemplo: ",{"type":32,"tag":126,"props":560,"children":562},{"className":561},[],[563],{"type":37,"value":564},"P(A > B) = 0.983",{"type":37,"value":566}," — A gana con 98.3% de confianza. El t-test frequentista con los mismos datos podría dar p=0.06 (no significativo), pero Bayesiano dice 98%. ¿Cuál es más relevante para decidir en negocio?",{"type":32,"tag":40,"props":568,"children":570},{"id":569},"pruebas-secuenciales-y-detención-temprana",[571],{"type":37,"value":572},"Pruebas Secuenciales y Detención Temprana",{"type":32,"tag":33,"props":574,"children":575},{},[576],{"type":37,"value":577},"El test Bayesiano está diseñado para ser secuencial. Actualiza el posterior cada día, verifica el umbral de decisión. Cuando \"Probability to be best\" supera 95%, detén la prueba e implementa el ganador. Esta detención temprana no infla el error Tipo I como en frequentista, porque el criterio de decisión es probabilidad posterior — no p-value.",{"type":32,"tag":33,"props":579,"children":580},{},[581],{"type":37,"value":582},"Implementación práctica:",{"type":32,"tag":584,"props":585,"children":586},"ol",{},[587,593,598,603,608,613],{"type":32,"tag":588,"props":589,"children":590},"li",{},[591],{"type":37,"value":592},"Define el prior (típicamente Beta(1,1) uniforme)",{"type":32,"tag":588,"props":594,"children":595},{},[596],{"type":37,"value":597},"Acumula datos de conversión diarios",{"type":32,"tag":588,"props":599,"children":600},{},[601],{"type":37,"value":602},"Calcula el posterior",{"type":32,"tag":588,"props":604,"children":605},{},[606],{"type":37,"value":607},"Calcula P(A > B) y P(B > A)",{"type":32,"tag":588,"props":609,"children":610},{},[611],{"type":37,"value":612},"Si cualquiera supera 95%, detén la prueba",{"type":32,"tag":588,"props":614,"children":615},{},[616],{"type":37,"value":617},"Si después de 14 días no alcanzas 95%, finaliza como \"no concluyente\" (datos insuficientes)",{"type":32,"tag":33,"props":619,"children":620},{},[621,623,632],{"type":37,"value":622},"Este enfoque es crítico en procesos de ",{"type":32,"tag":624,"props":625,"children":629},"a",{"href":626,"rel":627},"https:\u002F\u002Fwww.roibase.com.tr\u002Fes\u002Fcro",[628],"nofollow",[630],{"type":37,"value":631},"optimización de tasa de conversión",{"type":37,"value":633},". En una prueba de landing page donde la variante B muestra 30% menor CTR en CTA durante los primeros 3 días, el posterior Bayesiano dice \"96% B es peor\". La regla frequentist de tamaño muestral te obligaría a esperar 10 días más, pero tú detienes en el día 3, redirige tráfico a A. Menor costo de oportunidad.",{"type":32,"tag":112,"props":635,"children":637},{"id":636},"dinámicas-de-tamaño-muestral",[638],{"type":37,"value":639},"Dinámicas de Tamaño Muestral",{"type":32,"tag":33,"props":641,"children":642},{},[643],{"type":37,"value":644},"Bayesiano no requiere tamaño muestral fijo, pero puedes estimar el \"tamaño muestral esperado\". Depende de cuán informativo sea el prior. Si la tasa de conversión histórica es 10%, informas el prior con Beta(10,90) y necesitas menos datos. Con prior no-informativo tardará más, pero aún más rápido que frequentist.",{"type":32,"tag":33,"props":646,"children":647},{},[648],{"type":37,"value":649},"Tabla de simulación (ejemplo):",{"type":32,"tag":651,"props":652,"children":653},"table",{},[654,683],{"type":32,"tag":655,"props":656,"children":657},"thead",{},[658],{"type":32,"tag":659,"props":660,"children":661},"tr",{},[662,668,673,678],{"type":32,"tag":663,"props":664,"children":665},"th",{},[666],{"type":37,"value":667},"Verdadero Δ",{"type":32,"tag":663,"props":669,"children":670},{},[671],{"type":37,"value":672},"N Frequentista",{"type":32,"tag":663,"props":674,"children":675},{},[676],{"type":37,"value":677},"Expected N Bayesiano",{"type":32,"tag":663,"props":679,"children":680},{},[681],{"type":37,"value":682},"N Bayesiano percentil 90",{"type":32,"tag":684,"props":685,"children":686},"tbody",{},[687,711,734],{"type":32,"tag":659,"props":688,"children":689},{},[690,696,701,706],{"type":32,"tag":691,"props":692,"children":693},"td",{},[694],{"type":37,"value":695},"+10%",{"type":32,"tag":691,"props":697,"children":698},{},[699],{"type":37,"value":700},"4,800",{"type":32,"tag":691,"props":702,"children":703},{},[704],{"type":37,"value":705},"3,200",{"type":32,"tag":691,"props":707,"children":708},{},[709],{"type":37,"value":710},"5,100",{"type":32,"tag":659,"props":712,"children":713},{},[714,719,724,729],{"type":32,"tag":691,"props":715,"children":716},{},[717],{"type":37,"value":718},"+20%",{"type":32,"tag":691,"props":720,"children":721},{},[722],{"type":37,"value":723},"1,200",{"type":32,"tag":691,"props":725,"children":726},{},[727],{"type":37,"value":728},"800",{"type":32,"tag":691,"props":730,"children":731},{},[732],{"type":37,"value":733},"1,400",{"type":32,"tag":659,"props":735,"children":736},{},[737,742,747,752],{"type":32,"tag":691,"props":738,"children":739},{},[740],{"type":37,"value":741},"+5%",{"type":32,"tag":691,"props":743,"children":744},{},[745],{"type":37,"value":746},"19,200",{"type":32,"tag":691,"props":748,"children":749},{},[750],{"type":37,"value":751},"14,000",{"type":32,"tag":691,"props":753,"children":754},{},[755],{"type":37,"value":756},"22,000",{"type":32,"tag":33,"props":758,"children":759},{},[760],{"type":37,"value":761},"En lifts pequeños, Bayesiano también tarda pero no es tan rígido. En lifts grandes, 30-40% más rápido.",{"type":32,"tag":40,"props":763,"children":765},{"id":764},"contraargumentos-y-tradeoffs",[766],{"type":37,"value":767},"Contraargumentos y Tradeoffs",{"type":32,"tag":33,"props":769,"children":770},{},[771,776],{"type":32,"tag":73,"props":772,"children":773},{},[774],{"type":37,"value":775},"1. La elección del prior es subjetiva:",{"type":37,"value":777}," Cierto, introduces creencia previa. Pero con prior no-informativo (Beta(1,1)) minimizas este sesgo. Con suficientes datos, el likelihood domina y el prior se diluye. Frequentista parece \"objetivo\" pero las elecciones de alpha, power y MDE también son subjetivas.",{"type":32,"tag":33,"props":779,"children":780},{},[781,786],{"type":32,"tag":73,"props":782,"children":783},{},[784],{"type":37,"value":785},"2. Costo computacional:",{"type":37,"value":787}," Test Bayesiano requiere actualización posterior diaria + muestreo Monte Carlo. T-test frequentista es cálculo único. Pero las herramientas modernas (pymc, Stan, Google Optimize Bayesiano) lo automatizan. Extraer 10.000 muestras toma milisegundos — no es obstáculo.",{"type":32,"tag":33,"props":789,"children":790},{},[791,796],{"type":32,"tag":73,"props":792,"children":793},{},[794],{"type":37,"value":795},"3. Conformidad regulatoria:",{"type":37,"value":797}," En ensayos farmacéuticos con aprobación FDA, frequentist es estándar. En marketing digital, no hay restricción. Plataformas como Optimizely, VWO y AB Tasty ofrecen opciones Bayesianas.",{"type":32,"tag":33,"props":799,"children":800},{},[801,806],{"type":32,"tag":73,"props":802,"children":803},{},[804],{"type":37,"value":805},"4. Confusión con multi-armed bandits:",{"type":37,"value":807}," Prueba Bayesiana y algoritmos bandit (Thompson sampling) se confunden. Los bandits optimizan exploración-explotación, asignando más tráfico a variantes ganadoras durante la prueba. El test Bayesiano usa split fijo y usa posterior para decidir. Son casos de uso diferentes — bandit para campañas de alta velocidad, Bayesiano para cambios de producto de ciclo largo.",{"type":32,"tag":40,"props":809,"children":811},{"id":810},"escenario-real-prueba-de-creative-en-meta-ads",[812],{"type":37,"value":813},"Escenario Real: Prueba de Creative en Meta Ads",{"type":32,"tag":33,"props":815,"children":816},{},[817],{"type":37,"value":818},"Pruebas 3 variantes de creative en Meta Ads (A, B, C). Presupuesto diario $500, CPA objetivo $25. Frequentist requiere 1,000 conversiones por creative (poder 80%, MDE 15%). Con 60 conversiones diarias, esperas 50 días. Pero en el día 10, el CPA de C sube a $40 — obviamente malo.",{"type":32,"tag":33,"props":820,"children":821},{},[822],{"type":37,"value":823},"Con Bayesiano:",{"type":32,"tag":825,"props":826,"children":827},"ul",{},[828,833,838,843],{"type":32,"tag":588,"props":829,"children":830},{},[831],{"type":37,"value":832},"Acumula diario: spend, conversiones por creative",{"type":32,"tag":588,"props":834,"children":835},{},[836],{"type":37,"value":837},"Calcula distribución posterior de CPA (usa likelihood Gamma, CPA es positivo continuo)",{"type":32,"tag":588,"props":839,"children":840},{},[841],{"type":37,"value":842},"Calcula P(CPA_C > $30) = 92%",{"type":32,"tag":588,"props":844,"children":845},{},[846],{"type":37,"value":847},"Pausa C en el día 10, redistribuye presupuesto a A y B",{"type":32,"tag":33,"props":849,"children":850},{},[851],{"type":37,"value":852},"En el día 20, P(CPA_A \u003C CPA_B) = 96%. Declares A ganador. Decidiste en 20 días en lugar de 30. Ahorras $5,000 + 10 días ejecutando CPA mejor.",{"type":32,"tag":33,"props":854,"children":855},{},[856],{"type":37,"value":857},"Este tipo de decisión dinámica es crítica post-iOS14. La pérdida de señal debilitó la confiabilidad de pruebas — el posterior Bayesiano muestra incertidumbre explícitamente. \"Los datos son insuficientes, la distribución es muy ancha\" — el p-value frequentista no comunica esto.",{"type":32,"tag":859,"props":860,"children":861},"hr",{},[],{"type":32,"tag":33,"props":863,"children":864},{},[865],{"type":37,"value":866},"La prueba A\u002FB Bayesiana resuelve los problemas de rigidez de tamaño muestral y restricción de \"peeking\" del enfoque frequentista. Con testing secuencial, mides poder de decisión diario y detiene cuando alcanzan confianza suficiente — antes. La elección del prior introduce subjetividad pero prior no-informativo + abundancia de datos lo mitigua. En performance marketing, si necesitas flexibilidad de campaña, eficiencia presupuestaria y velocidad de decisión, el marco Bayesiano es correcto. Construye tu infraestructura de prueba alrededor de actualización posterior dinámica, no cálculo estático de N.",{"type":32,"tag":868,"props":869,"children":870},"style",{},[871],{"type":37,"value":872},"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":16,"searchDepth":182,"depth":182,"links":874},[875,876,879,882,883],{"id":42,"depth":159,"text":45},{"id":63,"depth":159,"text":66,"children":877},[878],{"id":114,"depth":182,"text":117},{"id":569,"depth":159,"text":572,"children":880},[881],{"id":636,"depth":182,"text":639},{"id":764,"depth":159,"text":767},{"id":810,"depth":159,"text":813},"markdown","content:es:marketing:prueba-bayesiana-ab-toma-rapida-decisiones.md","content","es\u002Fmarketing\u002Fprueba-bayesiana-ab-toma-rapida-decisiones.md","es\u002Fmarketing\u002Fprueba-bayesiana-ab-toma-rapida-decisiones","md",1782079490228]