[{"data":1,"prerenderedAt":1172},["ShallowReactive",2],{"article-alternates":3,"article-\u002Ffr\u002Fmarketing\u002Ftest-karar-bayesian":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":9,"_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":1166,"_id":1167,"_source":1168,"_file":1169,"_stem":1170,"_extension":1171},"marketing",false,"","Décisions Plus Rapides avec les Tests A\u002FB Bayésiens","Dépassez les pièges des tests fréquentistes avec la taille d'échantillon fixe. L'approche bayésienne, le monitoring séquentiel et l'arrêt précoce raccourcissent les processus de test de 40 à 60%.","2026-05-30",[20,21,22,23,24],"ab-testing","bayesian-statistics","experimentation","conversion-optimization","statistical-inference",8,"Roibase",{"type":28,"children":29,"toc":1155},"root",[30,60,67,79,104,132,154,170,183,200,228,238,262,278,300,317,324,808,814,825,835,859,875,880,896,909,920,940,960,989,1021,1048,1065,1071,1082,1098,1108,1124,1140,1144,1149],{"type":31,"tag":32,"props":33,"children":34},"element","p",{},[35,38,44,46,51,53,58],{"type":36,"value":37},"text","Dans le marketing de la performance, les tests A\u002FB forment l'épine dorsale de la prise de décision basée sur les preuves plutôt que sur l'hypothèse. Cependant, la plupart des équipes restent bloquées sur le dogme du ",{"type":31,"tag":39,"props":40,"children":41},"em",{},[42],{"type":36,"value":43},"sample size",{"type":36,"value":45}," fixe du paradigme fréquentiste : « N'observez pas jusqu'à atteindre le nombre calculé, car regarder trop tôt crée un biais. » Cette approche prolonge inutilement les processus de test sur 3 à 4 semaines. Le test A\u002FB bayésien permet le monitoring séquentiel avec la probabilité ",{"type":31,"tag":39,"props":47,"children":48},{},[49],{"type":36,"value":50},"posterior",{"type":36,"value":52},". Vous lisez les données quotidiennement, les fusionnez avec votre connaissance préalable (",{"type":31,"tag":39,"props":54,"children":55},{},[56],{"type":36,"value":57},"prior",{"type":36,"value":59},"), et arrêtez le test lorsque vous atteignez un seuil de confiance défini — par exemple, 95 % de probabilité d'être le meilleur. Résultat : prendre des décisions avec le même niveau de rigueur statistique, 40 à 60 % plus rapidement.",{"type":31,"tag":61,"props":62,"children":64},"h2",{"id":63},"les-limites-structurelles-de-lapproche-fréquentiste",[65],{"type":36,"value":66},"Les Limites Structurelles de l'Approche Fréquentiste",{"type":31,"tag":32,"props":68,"children":69},{},[70,72,77],{"type":36,"value":71},"Le test A\u002FB fréquentiste repose sur la ",{"type":31,"tag":39,"props":73,"children":74},{},[75],{"type":36,"value":76},"p-value",{"type":36,"value":78}," et l'intervalle de confiance. Vous testez la significativité de l'hypothèse nulle — tentant de rejeter l'hypothèse « il n'y a pas de différence entre A et B ». Les problèmes fondamentaux de cette approche :",{"type":31,"tag":32,"props":80,"children":81},{},[82,88,90,95,97,102],{"type":31,"tag":83,"props":84,"children":85},"strong",{},[86],{"type":36,"value":87},"Obligation d'une taille d'échantillon fixe.",{"type":36,"value":89}," Vous effectuez une analyse de puissance : taux de conversion baseline 2 %, effet minimum détectable (",{"type":31,"tag":39,"props":91,"children":92},{},[93],{"type":36,"value":94},"MDE",{"type":36,"value":96},") 10 % de levée relative, alpha 0,05, puissance 0,80. Vous devez mener le test jusqu'à atteindre la taille d'échantillon calculée (par exemple, 15 000 vues par variante). Si vous arrêtez tôt après avoir regardé, le problème des ",{"type":31,"tag":39,"props":98,"children":99},{},[100],{"type":36,"value":101},"comparaisons multiples",{"type":36,"value":103}," entre en jeu — le taux de faux positif dépasse la valeur alpha (0,05). En pratique : vous observez 25 % de levée au jour 2, mais attendez 3 semaines supplémentaires car « les données sont insuffisantes ».",{"type":31,"tag":32,"props":105,"children":106},{},[107,118,120,124,126,130],{"type":31,"tag":83,"props":108,"children":109},{},[110,112,116],{"type":36,"value":111},"Incapacité à exprimer l'incertitude ",{"type":31,"tag":39,"props":113,"children":114},{},[115],{"type":36,"value":50},{"type":36,"value":117},".",{"type":36,"value":119}," La ",{"type":31,"tag":39,"props":121,"children":122},{},[123],{"type":36,"value":76},{"type":36,"value":125}," vous indique « la probabilité d'observer un résultat aussi extrême ou plus sous l'hypothèse nulle ». Mais ce que vous voulez vraiment : « Quelle est la probabilité que la variante B soit réellement meilleure ? » Le cadre fréquentiste ne répond pas directement à cette question — ",{"type":31,"tag":39,"props":127,"children":128},{},[129],{"type":36,"value":32},{"type":36,"value":131}," \u003C 0,05 indique simplement qu'on rejette le null, sans exprimer la probabilité que B soit supérieure sous forme d'une probabilité.",{"type":31,"tag":32,"props":133,"children":134},{},[135,140,142,146,148,152],{"type":31,"tag":83,"props":136,"children":137},{},[138],{"type":36,"value":139},"Mécanisme de décision binaire.",{"type":36,"value":141}," Une ",{"type":31,"tag":39,"props":143,"children":144},{},[145],{"type":36,"value":76},{"type":36,"value":147}," de 0,049 signifie « significatif », 0,051 « non significatif ». Dans le monde réel, l'incertitude ne se divise pas aussi nettement. Vous ne pouvez pas interpréter une ",{"type":31,"tag":39,"props":149,"children":150},{},[151],{"type":36,"value":76},{"type":36,"value":153}," de 0,06 comme « il y a des preuves marginales, continuez le test » — c'est binaire, rejeter ou accepter.",{"type":31,"tag":32,"props":155,"children":156},{},[157,159,168],{"type":36,"value":158},"Ces limites structurelles ralentissent la vélocité des tests, en particulier dans les processus d'",{"type":31,"tag":160,"props":161,"children":165},"a",{"href":162,"rel":163},"https:\u002F\u002Fwww.roibase.com.tr\u002Ffr\u002Fcro",[164],"nofollow",[166],{"type":36,"value":167},"Optimisation du Taux de Conversion",{"type":36,"value":169},". Au lieu d'exécuter 2 à 3 itérations d'hypothèses par semaine, vous êtes bloqué par les règles de taille d'échantillon.",{"type":31,"tag":61,"props":171,"children":173},{"id":172},"test-bayésien-probabilité-posterior-et-monitoring-séquentiel",[174,176,181],{"type":36,"value":175},"Test Bayésien : Probabilité ",{"type":31,"tag":39,"props":177,"children":178},{},[179],{"type":36,"value":180},"Posterior",{"type":36,"value":182}," et Monitoring Séquentiel",{"type":31,"tag":32,"props":184,"children":185},{},[186,188,192,194,198],{"type":36,"value":187},"L'approche bayésienne traite le paramètre (taux de conversion) non pas comme un nombre fixe, mais comme une distribution de probabilité. La croyance ",{"type":31,"tag":39,"props":189,"children":190},{},[191],{"type":36,"value":57},{"type":36,"value":193}," (connaissance préalable) + données observées → distribution ",{"type":31,"tag":39,"props":195,"children":196},{},[197],{"type":36,"value":50},{"type":36,"value":199}," (conviction mise à jour). Formellement :",{"type":31,"tag":32,"props":201,"children":202},{},[203,214,216,220,222,226],{"type":31,"tag":83,"props":204,"children":205},{},[206,208,212],{"type":36,"value":207},"Distribution ",{"type":31,"tag":39,"props":209,"children":210},{},[211],{"type":36,"value":57},{"type":36,"value":213}," :",{"type":36,"value":215}," Votre conviction préalable sur le taux de conversion baseline. Si vous n'avez aucune connaissance, vous utilisez un ",{"type":31,"tag":39,"props":217,"children":218},{},[219],{"type":36,"value":57},{"type":36,"value":221}," non-informatif (Beta(1,1)) — égale probabilité pour tous les paramètres. Si vous savez que « le taux de conversion se situe généralement entre 1,5 % et 2,5 % » à partir de tests précédents, vous définissez un ",{"type":31,"tag":39,"props":223,"children":224},{},[225],{"type":36,"value":57},{"type":36,"value":227}," informatif (Beta(15, 985)).",{"type":31,"tag":32,"props":229,"children":230},{},[231,236],{"type":31,"tag":83,"props":232,"children":233},{},[234],{"type":36,"value":235},"Vraisemblance :",{"type":36,"value":237}," Les données observées — 1 000 vues, 25 conversions, par exemple.",{"type":31,"tag":32,"props":239,"children":240},{},[241,246,248,252,254,261],{"type":31,"tag":83,"props":242,"children":243},{},[244],{"type":36,"value":245},"Posterior :",{"type":36,"value":247}," Distribution mise à jour via le théorème de Bayes. Utilisant une paire conjuguée Beta-binomiale, le ",{"type":31,"tag":39,"props":249,"children":250},{},[251],{"type":36,"value":50},{"type":36,"value":253}," se résout analytiquement : ",{"type":31,"tag":255,"props":256,"children":258},"code",{"className":257},[],[259],{"type":36,"value":260},"Beta(alpha + conversions, beta + non_conversions)",{"type":36,"value":117},{"type":31,"tag":32,"props":263,"children":264},{},[265,270,272,276],{"type":31,"tag":83,"props":266,"children":267},{},[268],{"type":36,"value":269},"Règle de décision :",{"type":36,"value":271}," Vous échantillonnez les distributions ",{"type":31,"tag":39,"props":273,"children":274},{},[275],{"type":36,"value":50},{"type":36,"value":277}," des variantes A et B via simulation Monte Carlo (par exemple, 100 000 itérations). À chaque itération, vous comptabilisez le nombre de fois où B dépasse A. Ce ratio est « la probabilité que B gagne » (P(B > A)). Si ce probabilité dépasse 95 %, vous arrêtez le test et choisissez B.",{"type":31,"tag":32,"props":279,"children":280},{},[281,286,288,292,294,298],{"type":31,"tag":83,"props":282,"children":283},{},[284],{"type":36,"value":285},"Monitoring séquentiel :",{"type":36,"value":287}," Le cadre bayésien vous permet de recalculer le ",{"type":31,"tag":39,"props":289,"children":290},{},[291],{"type":36,"value":50},{"type":36,"value":293}," quotidiennement. Contrairement au problème du « peeking » en approche fréquentiste, la mise à jour ",{"type":31,"tag":39,"props":295,"children":296},{},[297],{"type":36,"value":50},{"type":36,"value":299}," est une partie naturelle de l'inférence bayésienne. Chaque matin, quand vous ouvrez le tableau de bord, vous voyez la valeur actuelle de P(B > A) : 65 % → 78 % → 89 % → 94 % → 96 %. Une fois le seuil de 95 % atteint, arrêtez le test.",{"type":31,"tag":32,"props":301,"children":302},{},[303,305,309,311,315],{"type":36,"value":304},"En pratique : taux de conversion baseline 2 %, objectif 10 % de levée relative (soit 2,2 %), seuil de confiance 95 %. Le test fréquentiste demande 15 000 vues par variante (21 jours au total). Le test bayésien atteint le même seuil en 9 à 12 jours — car la connaissance ",{"type":31,"tag":39,"props":306,"children":307},{},[308],{"type":36,"value":57},{"type":36,"value":310}," permet un ",{"type":31,"tag":39,"props":312,"children":313},{},[314],{"type":36,"value":50},{"type":36,"value":316}," plus tranchant avec moins de données.",{"type":31,"tag":318,"props":319,"children":321},"h3",{"id":320},"exemple-de-code-de-simulation-python",[322],{"type":36,"value":323},"Exemple de Code de Simulation (Python)",{"type":31,"tag":325,"props":326,"children":330},"pre",{"className":327,"code":328,"language":329,"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# Données observées (jour 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# Sortie exemple : P(B > A) = 0.923 → en dessous de 95 %, continuer le test\n","python",[331],{"type":31,"tag":255,"props":332,"children":333},{"__ignoreMap":15},[334,362,385,395,405,435,460,468,476,503,529,537,546,593,638,646,655,683,709,737,745,799],{"type":31,"tag":335,"props":336,"children":339},"span",{"class":337,"line":338},"line",1,[340,346,352,357],{"type":31,"tag":335,"props":341,"children":343},{"style":342},"--shiki-default:#F97583",[344],{"type":36,"value":345},"import",{"type":31,"tag":335,"props":347,"children":349},{"style":348},"--shiki-default:#E1E4E8",[350],{"type":36,"value":351}," numpy ",{"type":31,"tag":335,"props":353,"children":354},{"style":342},[355],{"type":36,"value":356},"as",{"type":31,"tag":335,"props":358,"children":359},{"style":348},[360],{"type":36,"value":361}," np\n",{"type":31,"tag":335,"props":363,"children":365},{"class":337,"line":364},2,[366,371,376,380],{"type":31,"tag":335,"props":367,"children":368},{"style":342},[369],{"type":36,"value":370},"from",{"type":31,"tag":335,"props":372,"children":373},{"style":348},[374],{"type":36,"value":375}," scipy.stats ",{"type":31,"tag":335,"props":377,"children":378},{"style":342},[379],{"type":36,"value":345},{"type":31,"tag":335,"props":381,"children":382},{"style":348},[383],{"type":36,"value":384}," beta\n",{"type":31,"tag":335,"props":386,"children":388},{"class":337,"line":387},3,[389],{"type":31,"tag":335,"props":390,"children":392},{"emptyLinePlaceholder":391},true,[393],{"type":36,"value":394},"\n",{"type":31,"tag":335,"props":396,"children":398},{"class":337,"line":397},4,[399],{"type":31,"tag":335,"props":400,"children":402},{"style":401},"--shiki-default:#6A737D",[403],{"type":36,"value":404},"# Prior: Beta(1, 1) — uniform\n",{"type":31,"tag":335,"props":406,"children":408},{"class":337,"line":407},5,[409,414,419,425,430],{"type":31,"tag":335,"props":410,"children":411},{"style":348},[412],{"type":36,"value":413},"alpha_a, beta_a ",{"type":31,"tag":335,"props":415,"children":416},{"style":342},[417],{"type":36,"value":418},"=",{"type":31,"tag":335,"props":420,"children":422},{"style":421},"--shiki-default:#79B8FF",[423],{"type":36,"value":424}," 1",{"type":31,"tag":335,"props":426,"children":427},{"style":348},[428],{"type":36,"value":429},", ",{"type":31,"tag":335,"props":431,"children":432},{"style":421},[433],{"type":36,"value":434},"1\n",{"type":31,"tag":335,"props":436,"children":438},{"class":337,"line":437},6,[439,444,448,452,456],{"type":31,"tag":335,"props":440,"children":441},{"style":348},[442],{"type":36,"value":443},"alpha_b, beta_b ",{"type":31,"tag":335,"props":445,"children":446},{"style":342},[447],{"type":36,"value":418},{"type":31,"tag":335,"props":449,"children":450},{"style":421},[451],{"type":36,"value":424},{"type":31,"tag":335,"props":453,"children":454},{"style":348},[455],{"type":36,"value":429},{"type":31,"tag":335,"props":457,"children":458},{"style":421},[459],{"type":36,"value":434},{"type":31,"tag":335,"props":461,"children":463},{"class":337,"line":462},7,[464],{"type":31,"tag":335,"props":465,"children":466},{"emptyLinePlaceholder":391},[467],{"type":36,"value":394},{"type":31,"tag":335,"props":469,"children":470},{"class":337,"line":25},[471],{"type":31,"tag":335,"props":472,"children":473},{"style":401},[474],{"type":36,"value":475},"# Données observées (jour 5)\n",{"type":31,"tag":335,"props":477,"children":479},{"class":337,"line":478},9,[480,485,489,494,498],{"type":31,"tag":335,"props":481,"children":482},{"style":348},[483],{"type":36,"value":484},"views_a, conv_a ",{"type":31,"tag":335,"props":486,"children":487},{"style":342},[488],{"type":36,"value":418},{"type":31,"tag":335,"props":490,"children":491},{"style":421},[492],{"type":36,"value":493}," 5000",{"type":31,"tag":335,"props":495,"children":496},{"style":348},[497],{"type":36,"value":429},{"type":31,"tag":335,"props":499,"children":500},{"style":421},[501],{"type":36,"value":502},"95\n",{"type":31,"tag":335,"props":504,"children":506},{"class":337,"line":505},10,[507,512,516,520,524],{"type":31,"tag":335,"props":508,"children":509},{"style":348},[510],{"type":36,"value":511},"views_b, conv_b ",{"type":31,"tag":335,"props":513,"children":514},{"style":342},[515],{"type":36,"value":418},{"type":31,"tag":335,"props":517,"children":518},{"style":421},[519],{"type":36,"value":493},{"type":31,"tag":335,"props":521,"children":522},{"style":348},[523],{"type":36,"value":429},{"type":31,"tag":335,"props":525,"children":526},{"style":421},[527],{"type":36,"value":528},"112\n",{"type":31,"tag":335,"props":530,"children":532},{"class":337,"line":531},11,[533],{"type":31,"tag":335,"props":534,"children":535},{"emptyLinePlaceholder":391},[536],{"type":36,"value":394},{"type":31,"tag":335,"props":538,"children":540},{"class":337,"line":539},12,[541],{"type":31,"tag":335,"props":542,"children":543},{"style":401},[544],{"type":36,"value":545},"# Posterior\n",{"type":31,"tag":335,"props":547,"children":549},{"class":337,"line":548},13,[550,555,559,564,569,574,578,583,588],{"type":31,"tag":335,"props":551,"children":552},{"style":348},[553],{"type":36,"value":554},"post_a ",{"type":31,"tag":335,"props":556,"children":557},{"style":342},[558],{"type":36,"value":418},{"type":31,"tag":335,"props":560,"children":561},{"style":348},[562],{"type":36,"value":563}," beta(alpha_a ",{"type":31,"tag":335,"props":565,"children":566},{"style":342},[567],{"type":36,"value":568},"+",{"type":31,"tag":335,"props":570,"children":571},{"style":348},[572],{"type":36,"value":573}," conv_a, beta_a ",{"type":31,"tag":335,"props":575,"children":576},{"style":342},[577],{"type":36,"value":568},{"type":31,"tag":335,"props":579,"children":580},{"style":348},[581],{"type":36,"value":582}," views_a ",{"type":31,"tag":335,"props":584,"children":585},{"style":342},[586],{"type":36,"value":587},"-",{"type":31,"tag":335,"props":589,"children":590},{"style":348},[591],{"type":36,"value":592}," conv_a)\n",{"type":31,"tag":335,"props":594,"children":596},{"class":337,"line":595},14,[597,602,606,611,615,620,624,629,633],{"type":31,"tag":335,"props":598,"children":599},{"style":348},[600],{"type":36,"value":601},"post_b ",{"type":31,"tag":335,"props":603,"children":604},{"style":342},[605],{"type":36,"value":418},{"type":31,"tag":335,"props":607,"children":608},{"style":348},[609],{"type":36,"value":610}," beta(alpha_b ",{"type":31,"tag":335,"props":612,"children":613},{"style":342},[614],{"type":36,"value":568},{"type":31,"tag":335,"props":616,"children":617},{"style":348},[618],{"type":36,"value":619}," conv_b, beta_b ",{"type":31,"tag":335,"props":621,"children":622},{"style":342},[623],{"type":36,"value":568},{"type":31,"tag":335,"props":625,"children":626},{"style":348},[627],{"type":36,"value":628}," views_b ",{"type":31,"tag":335,"props":630,"children":631},{"style":342},[632],{"type":36,"value":587},{"type":31,"tag":335,"props":634,"children":635},{"style":348},[636],{"type":36,"value":637}," conv_b)\n",{"type":31,"tag":335,"props":639,"children":641},{"class":337,"line":640},15,[642],{"type":31,"tag":335,"props":643,"children":644},{"emptyLinePlaceholder":391},[645],{"type":36,"value":394},{"type":31,"tag":335,"props":647,"children":649},{"class":337,"line":648},16,[650],{"type":31,"tag":335,"props":651,"children":652},{"style":401},[653],{"type":36,"value":654},"# Monte Carlo: P(B > A)\n",{"type":31,"tag":335,"props":656,"children":658},{"class":337,"line":657},17,[659,664,668,673,678],{"type":31,"tag":335,"props":660,"children":661},{"style":348},[662],{"type":36,"value":663},"samples_a ",{"type":31,"tag":335,"props":665,"children":666},{"style":342},[667],{"type":36,"value":418},{"type":31,"tag":335,"props":669,"children":670},{"style":348},[671],{"type":36,"value":672}," post_a.rvs(",{"type":31,"tag":335,"props":674,"children":675},{"style":421},[676],{"type":36,"value":677},"100000",{"type":31,"tag":335,"props":679,"children":680},{"style":348},[681],{"type":36,"value":682},")\n",{"type":31,"tag":335,"props":684,"children":686},{"class":337,"line":685},18,[687,692,696,701,705],{"type":31,"tag":335,"props":688,"children":689},{"style":348},[690],{"type":36,"value":691},"samples_b ",{"type":31,"tag":335,"props":693,"children":694},{"style":342},[695],{"type":36,"value":418},{"type":31,"tag":335,"props":697,"children":698},{"style":348},[699],{"type":36,"value":700}," post_b.rvs(",{"type":31,"tag":335,"props":702,"children":703},{"style":421},[704],{"type":36,"value":677},{"type":31,"tag":335,"props":706,"children":707},{"style":348},[708],{"type":36,"value":682},{"type":31,"tag":335,"props":710,"children":712},{"class":337,"line":711},19,[713,718,722,727,732],{"type":31,"tag":335,"props":714,"children":715},{"style":348},[716],{"type":36,"value":717},"prob_b_wins ",{"type":31,"tag":335,"props":719,"children":720},{"style":342},[721],{"type":36,"value":418},{"type":31,"tag":335,"props":723,"children":724},{"style":348},[725],{"type":36,"value":726}," (samples_b ",{"type":31,"tag":335,"props":728,"children":729},{"style":342},[730],{"type":36,"value":731},">",{"type":31,"tag":335,"props":733,"children":734},{"style":348},[735],{"type":36,"value":736}," samples_a).mean()\n",{"type":31,"tag":335,"props":738,"children":740},{"class":337,"line":739},20,[741],{"type":31,"tag":335,"props":742,"children":743},{"emptyLinePlaceholder":391},[744],{"type":36,"value":394},{"type":31,"tag":335,"props":746,"children":748},{"class":337,"line":747},21,[749,754,759,764,770,775,780,785,790,795],{"type":31,"tag":335,"props":750,"children":751},{"style":421},[752],{"type":36,"value":753},"print",{"type":31,"tag":335,"props":755,"children":756},{"style":348},[757],{"type":36,"value":758},"(",{"type":31,"tag":335,"props":760,"children":761},{"style":342},[762],{"type":36,"value":763},"f",{"type":31,"tag":335,"props":765,"children":767},{"style":766},"--shiki-default:#9ECBFF",[768],{"type":36,"value":769},"\"P(B > A) = ",{"type":31,"tag":335,"props":771,"children":772},{"style":421},[773],{"type":36,"value":774},"{",{"type":31,"tag":335,"props":776,"children":777},{"style":348},[778],{"type":36,"value":779},"prob_b_wins",{"type":31,"tag":335,"props":781,"children":782},{"style":342},[783],{"type":36,"value":784},":.3f",{"type":31,"tag":335,"props":786,"children":787},{"style":421},[788],{"type":36,"value":789},"}",{"type":31,"tag":335,"props":791,"children":792},{"style":766},[793],{"type":36,"value":794},"\"",{"type":31,"tag":335,"props":796,"children":797},{"style":348},[798],{"type":36,"value":682},{"type":31,"tag":335,"props":800,"children":802},{"class":337,"line":801},22,[803],{"type":31,"tag":335,"props":804,"children":805},{"style":401},[806],{"type":36,"value":807},"# Sortie exemple : P(B > A) = 0.923 → en dessous de 95 %, continuer le test\n",{"type":31,"tag":61,"props":809,"children":811},{"id":810},"dynamique-de-la-taille-déchantillon-et-critères-darrêt-précoce",[812],{"type":36,"value":813},"Dynamique de la Taille d'Échantillon et Critères d'Arrêt Précoce",{"type":31,"tag":32,"props":815,"children":816},{},[817,819,823],{"type":36,"value":818},"L'avantage de rapidité du test bayésien provient de la taille d'échantillon dynamique. Au lieu de viser un N fixe, vous liez la règle d'arrêt à la confiance ",{"type":31,"tag":39,"props":820,"children":821},{},[822],{"type":36,"value":50},{"type":36,"value":824},". Deux critères courants :",{"type":31,"tag":32,"props":826,"children":827},{},[828,833],{"type":31,"tag":83,"props":829,"children":830},{},[831],{"type":36,"value":832},"Seuil de probabilité :",{"type":36,"value":834}," P(B > A) ≥ 0,95, arrêt. Cela signifie « probabilité 95 % que B soit réellement meilleur ». Certaines équipes utilisent 99 % (plus conservateur), d'autres 90 % (plus agressif — pour la vélocité des tests).",{"type":31,"tag":32,"props":836,"children":837},{},[838,843,845,850,852,857],{"type":31,"tag":83,"props":839,"children":840},{},[841],{"type":36,"value":842},"Perte attendue :",{"type":36,"value":844}," Si vous choisissez B et qu'A s'avère réellement meilleur, quel est votre coût ? ",{"type":31,"tag":39,"props":846,"children":847},{},[848],{"type":36,"value":849},"Expected loss",{"type":36,"value":851}," = E",{"type":31,"tag":335,"props":853,"children":854},{},[855],{"type":36,"value":856},"max(0, A - B)",{"type":36,"value":858},". Si cette perte tombe en dessous d'un seuil acceptable (par exemple, \u003C 0,0001 différence de taux de conversion absolue), arrêtez le test. Cette métrique fournit une gestion des risques basée sur « le coût d'une mauvaise décision ».",{"type":31,"tag":32,"props":860,"children":861},{},[862,867,869,873],{"type":31,"tag":83,"props":863,"children":864},{},[865],{"type":36,"value":866},"Plancher d'échantillon minimum :",{"type":36,"value":868}," Pour éviter un arrêt complètement prématuré, posez une règle « collectez au minimum 3 000 échantillons, puis appliquez la règle d'arrêt bayésienne ». Cela empêche le ",{"type":31,"tag":39,"props":870,"children":871},{},[872],{"type":36,"value":57},{"type":36,"value":874}," de dominer excessivement.",{"type":31,"tag":32,"props":876,"children":877},{},[878],{"type":36,"value":879},"Scénario exemple : test de couleur d'appel à l'action sur un site d'e-commerce (vert vs. orange). Taux de conversion baseline 3,2 %. Semaine 1 : 8 000 vues, P(orange > vert) = 87 %. Semaine 2 : 16 000 vues, P = 94 %. Jour 2 de la semaine 3 (18 500 vues au total), P = 96 %. Le test fréquentiste aurait exigé 25 000 vues (18 jours au total), vous avez arrêté au jour 10. Vous avez réduit la durée du test de 44 %.",{"type":31,"tag":32,"props":881,"children":882},{},[883,888,890,894],{"type":31,"tag":39,"props":884,"children":885},{},[886],{"type":36,"value":887},"Tradeoff",{"type":36,"value":889}," : L'arrêt précoce risque, dans certains cas, de sélectionner une variante « chancheuse au démarrage mais en régression ». Pour atténuer : (1) imposez un plancher d'échantillon, (2) si l'effet est petit (par exemple, 5 % de levée relative), augmentez le seuil à 99 %, (3) surveillez l'écart-type du ",{"type":31,"tag":39,"props":891,"children":892},{},[893],{"type":36,"value":50},{"type":36,"value":895}," — s'il reste large (incertitude élevée), collectez plus de données.",{"type":31,"tag":61,"props":897,"children":899},{"id":898},"choix-du-prior-et-accumulation-de-connaissance",[900,902,907],{"type":36,"value":901},"Choix du ",{"type":31,"tag":39,"props":903,"children":904},{},[905],{"type":36,"value":906},"Prior",{"type":36,"value":908}," et Accumulation de Connaissance",{"type":31,"tag":32,"props":910,"children":911},{},[912,914,918],{"type":36,"value":913},"La puissance du test bayésien provient de la formalisation de la connaissance préalable. Mais un mauvais choix de ",{"type":31,"tag":39,"props":915,"children":916},{},[917],{"type":36,"value":57},{"type":36,"value":919}," peut créer un biais. Deux extrêmes :",{"type":31,"tag":32,"props":921,"children":922},{},[923,932,934,938],{"type":31,"tag":83,"props":924,"children":925},{},[926,930],{"type":31,"tag":39,"props":927,"children":928},{},[929],{"type":36,"value":906},{"type":36,"value":931}," non-informatif (Beta(1,1)) :",{"type":36,"value":933}," Aucune connaissance préalable. Chaque test recommence à zéro. Avantage : impartial. Inconvénient : le ",{"type":31,"tag":39,"props":935,"children":936},{},[937],{"type":36,"value":50},{"type":36,"value":939}," met plus de temps à devenir tranchant — taille d'échantillon similaire aux tests fréquentistes.",{"type":31,"tag":32,"props":941,"children":942},{},[943,952,954,958],{"type":31,"tag":83,"props":944,"children":945},{},[946,950],{"type":31,"tag":39,"props":947,"children":948},{},[949],{"type":36,"value":906},{"type":36,"value":951}," informatif (Beta(α, β)) :",{"type":36,"value":953}," Vous transportez la connaissance des tests précédents, des benchmarks sectoriels ou du baseline. Exemple : « Dans les tests de bouton CTA, le taux de conversion se situe généralement entre 2 % et 4 %, moyenne 2,8 % », vous définissez le ",{"type":31,"tag":39,"props":955,"children":956},{},[957],{"type":36,"value":57},{"type":36,"value":959}," Beta(28, 972) (moyenne 2,8 %, variance adaptée).",{"type":31,"tag":32,"props":961,"children":962},{},[963,965,969,971,975,977,981,983,987],{"type":36,"value":964},"Utiliser un ",{"type":31,"tag":39,"props":966,"children":967},{},[968],{"type":36,"value":57},{"type":36,"value":970}," informatif raccourcit les tests car ",{"type":31,"tag":39,"props":972,"children":973},{},[974],{"type":36,"value":57},{"type":36,"value":976}," + nouvelles données convergent plus rapidement. Mais risque : si le ",{"type":31,"tag":39,"props":978,"children":979},{},[980],{"type":36,"value":57},{"type":36,"value":982}," est incorrect (par exemple, copié d'une vertical ancienne, le nouveau segment diffère), le ",{"type":31,"tag":39,"props":984,"children":985},{},[986],{"type":36,"value":50},{"type":36,"value":988}," est biaisé. Deux protections :",{"type":31,"tag":32,"props":990,"children":991},{},[992,1002,1004,1008,1010,1014,1016,1020],{"type":31,"tag":83,"props":993,"children":994},{},[995,997,1001],{"type":36,"value":996},"Analyse de sensibilité du ",{"type":31,"tag":39,"props":998,"children":999},{},[1000],{"type":36,"value":57},{"type":36,"value":213},{"type":36,"value":1003}," Exécutez le test avec différents ",{"type":31,"tag":39,"props":1005,"children":1006},{},[1007],{"type":36,"value":57},{"type":36,"value":1009}," (faible, moyen, fort informatif) et vérifiez si les résultats changent. Si le résultat varie énormément selon le ",{"type":31,"tag":39,"props":1011,"children":1012},{},[1013],{"type":36,"value":57},{"type":36,"value":1015}," (par exemple, 60 % vs. 98 % de probabilité de gagne), allongez le test — les données ne surpassent pas encore le ",{"type":31,"tag":39,"props":1017,"children":1018},{},[1019],{"type":36,"value":57},{"type":36,"value":117},{"type":31,"tag":32,"props":1022,"children":1023},{},[1024,1033,1035,1039,1041,1046],{"type":31,"tag":83,"props":1025,"children":1026},{},[1027,1031],{"type":31,"tag":39,"props":1028,"children":1029},{},[1030],{"type":36,"value":906},{"type":36,"value":1032}," hiérarchique :",{"type":36,"value":1034}," Si vous testez sur plusieurs segments (mobile vs. desktop, par pays), utilisez un modèle bayésien hiérarchique. Chaque segment a son propre taux de conversion, mais tous sont rétrécis depuis une moyenne globale ",{"type":31,"tag":39,"props":1036,"children":1037},{},[1038],{"type":36,"value":57},{"type":36,"value":1040},". Cela réduit le ",{"type":31,"tag":39,"props":1042,"children":1043},{},[1044],{"type":36,"value":1045},"overfitting",{"type":36,"value":1047}," au niveau du segment.",{"type":31,"tag":32,"props":1049,"children":1050},{},[1051,1053,1057,1059,1063],{"type":36,"value":1052},"Recommandation pratique : exécutez les 5 à 10 premiers tests avec un ",{"type":31,"tag":39,"props":1054,"children":1055},{},[1056],{"type":36,"value":57},{"type":36,"value":1058}," non-informatif, accumulez les résultats, calculez la moyenne et la variance, puis utilisez cette information comme ",{"type":31,"tag":39,"props":1060,"children":1061},{},[1062],{"type":36,"value":57},{"type":36,"value":1064}," informatif pour les tests suivants. Cette approche « meta-learning » grave la connaissance cumulative des tests dans la mémoire.",{"type":31,"tag":61,"props":1066,"children":1068},{"id":1067},"intégration-organisationnelle-et-protocole-de-décision",[1069],{"type":36,"value":1070},"Intégration Organisationnelle et Protocole de Décision",{"type":31,"tag":32,"props":1072,"children":1073},{},[1074,1076,1080],{"type":36,"value":1075},"Intégrer le test A\u002FB bayésien à la culture d'équipe est une question organisationnelle, non technique. Quand vous dites à une équipe habituée aux tests fréquentistes « vous pouvez maintenant consulter chaque jour », la réaction initiale est mélangée : « Où est la ",{"type":31,"tag":39,"props":1077,"children":1078},{},[1079],{"type":36,"value":76},{"type":36,"value":1081}," ? » Deux étapes :",{"type":31,"tag":32,"props":1083,"children":1084},{},[1085,1090,1092,1096],{"type":31,"tag":83,"props":1086,"children":1087},{},[1088],{"type":36,"value":1089},"Formation et onboarding :",{"type":36,"value":1091}," Expliquez ce que signifie la métrique P(B > A). Les équipes doivent communiquer à l'aise avec « 95 % de probabilité que B soit meilleur ». Au lieu de l'indirection fréquentiste « ",{"type":31,"tag":39,"props":1093,"children":1094},{},[1095],{"type":36,"value":32},{"type":36,"value":1097}," \u003C 0,05 donc null rejeté », vous avez un langage de décision direct. Exécutez les 2 à 3 premiers tests en parallèle — analysez avec l'approche fréquentiste et bayésienne, comparez. Quand l'équipe voit la différence, l'adoption s'accélère.",{"type":31,"tag":32,"props":1099,"children":1100},{},[1101,1106],{"type":31,"tag":83,"props":1102,"children":1103},{},[1104],{"type":36,"value":1105},"Standardisation du seuil de décision :",{"type":36,"value":1107}," À quel probabilité arrêtez-vous le test ? 95 % ? 99 % ? Cela dépend de la tolérance au risque. Trafic élevé + décisions à faible risque (par exemple, sujet d'email) → 90 % suffit. Trafic faible + décisions coûteuses (par exemple, refonte page tarification) → utilisez 99 %. Documentez ces seuils dans votre playbook de tests.",{"type":31,"tag":32,"props":1109,"children":1110},{},[1111,1116,1118,1122],{"type":31,"tag":83,"props":1112,"children":1113},{},[1114],{"type":36,"value":1115},"Monitoring post-test :",{"type":36,"value":1117}," Vous arrêtez le test, déclarez B gagnant. Mais deux semaines après rollout complet, le taux de conversion chute — régression vers la moyenne ou facteur externe (campagne, saisonnalité). Le test bayésien réduit ce risque mais ne l'élimine pas. Solution : monitorer 1 semaine post-rollout, si la moyenne ",{"type":31,"tag":39,"props":1119,"children":1120},{},[1121],{"type":36,"value":50},{"type":36,"value":1123}," chute > 10 %, déclenchez un rollback.",{"type":31,"tag":32,"props":1125,"children":1126},{},[1127,1132,1134,1138],{"type":31,"tag":83,"props":1128,"children":1129},{},[1130],{"type":36,"value":1131},"Outillage :",{"type":36,"value":1133}," Google Optimize offre un mode bayésien mais limité. VWO, Optimizely le supportent partiellement. Pour une stack personnalisée : Python (PyMC3, ArviZ) + BigQuery + tableau de bord Looker. Un job Airflow quotidien met à jour les ",{"type":31,"tag":39,"props":1135,"children":1136},{},[1137],{"type":36,"value":50},{"type":36,"value":1139},", Looker affiche la métrique P(B > A). Une alerte Slack se déclenche au dépassement du seuil.",{"type":31,"tag":1141,"props":1142,"children":1143},"hr",{},[],{"type":31,"tag":32,"props":1145,"children":1146},{},[1147],{"type":36,"value":1148},"Le test A\u002FB bayésien augmente la vélocité des tests mais exige une discipline statistique. Vous dé",{"type":31,"tag":1150,"props":1151,"children":1152},"style",{},[1153],{"type":36,"value":1154},"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":387,"depth":387,"links":1156},[1157,1158,1162,1163,1165],{"id":63,"depth":364,"text":66},{"id":172,"depth":364,"text":1159,"children":1160},"Test Bayésien : Probabilité Posterior et Monitoring Séquentiel",[1161],{"id":320,"depth":387,"text":323},{"id":810,"depth":364,"text":813},{"id":898,"depth":364,"text":1164},"Choix du Prior et Accumulation de Connaissance",{"id":1067,"depth":364,"text":1070},"markdown","content:fr:marketing:test-karar-bayesian.md","content","fr\u002Fmarketing\u002Ftest-karar-bayesian.md","fr\u002Fmarketing\u002Ftest-karar-bayesian","md",1781791483080]