[{"data":1,"prerenderedAt":952},["ShallowReactive",2],{"article-alternates":3,"article-\u002Ffr\u002Fgaming\u002Foptimisation-bayesienne-des-prix-f2p-mobile":13},{"i18nKey":4,"paths":5},"gaming-002-2026-08",{"de":6,"en":7,"es":8,"fr":9,"it":10,"ru":11,"tr":12},"\u002Fde\u002Fgaming\u002Fbayesian-price-optimization-mobile-f2p","\u002Fen\u002Fgaming\u002Fmobile-f2p-bayesian-price-optimization","\u002Fes\u002Fgaming\u002Fbayesian-precio-optimizacion-f2p-movil","\u002Ffr\u002Fgaming\u002Foptimisation-bayesienne-des-prix-f2p-mobile","\u002Fit\u002Fgaming\u002Fbayesian-price-optimization-f2p-mobile","\u002Fru\u002Fgaming\u002Fbayesian-cenovaya-optimizacija-v-mobile-f2p","\u002Ftr\u002Fgaming\u002Fmobile-f2pde-bayesian-price-optimization",{"_path":9,"_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":946,"_id":947,"_source":948,"_file":949,"_stem":950,"_extension":951},"gaming",false,"","Optimisation Bayésienne des Prix dans les F2P Mobile","Optimiser les paliers de prix des IAP avec des tests bayésiens : estimation posterior, tarification segmentée et méthodologie de calcul du lift revenue.","2026-08-05",[21,22,23,24,25],"monetisation-f2p","test-bayesien","optimisation-iap","price-ladder","mobile-gaming",9,"Roibase",{"type":29,"children":30,"toc":932},"root",[31,47,54,66,84,89,103,132,155,161,166,178,183,189,214,225,239,245,269,280,292,298,303,308,313,796,815,821,844,849,874,880,891,902,913,926],{"type":32,"tag":33,"props":34,"children":35},"element","p",{},[36,39,45],{"type":37,"value":38},"text","Dans les jeux F2P mobile, l'optimisation des prix IAP se réduit souvent à un simple test A\u002FB : comparez deux prix, choisissez celui qui génère le plus de revenue. Cette approche fonctionnait en 2018 — les coûts d'acquisition étaient bas et les problèmes de taille d'échantillon n'existaient pas. En 2026, la situation est différente : après iOS 14.5, le suivi des cohortes est fragmenté, le CPI Apple Search Ads a augmenté de 340 %, les durées de test sont passées de 8 à 14 semaines. La méthodologie bayésienne offre deux avantages majeurs dans ces conditions : elle permet de prendre des décisions précoces via la distribution posterior, et la segmentation renforce le modèle grâce aux ",{"type":32,"tag":40,"props":41,"children":42},"em",{},[43],{"type":37,"value":44},"priors",{"type":37,"value":46}," informatifs. En économie des jeux, l'élasticité des prix n'est pas constante — elle varie significativement entre les segments whale\u002Fdolphin\u002Fminnow, et capture cette différence dépasse les capacités des tests A\u002FB fréquentistes.",{"type":32,"tag":48,"props":49,"children":51},"h2",{"id":50},"la-logique-économique-des-tests-bayésiens",[52],{"type":37,"value":53},"La Logique Économique des Tests Bayésiens",{"type":32,"tag":33,"props":55,"children":56},{},[57,59,64],{"type":37,"value":58},"En F2P mobile, le coût d'un test de prix ne se limite pas au temps de développement : il inclut le ",{"type":32,"tag":40,"props":60,"children":61},{},[62],{"type":37,"value":63},"coût d'opportunité",{"type":37,"value":65},". Si vous testez un passage de $4.99 à $6.99 pendant 14 semaines, la revenue perdue en attendant le bon prix est elle-même un coût du test. L'approche bayésienne met à jour la distribution posterior chaque jour — au lieu d'une conversion de 2,3 %, vous avez un intervalle de crédibilité à 95 % : 1,8 % à 2,9 %. Cet intervalle se rétrécit progressivement, et lorsqu'il devient suffisamment étroit, la décision devient évidente et le test peut s'arrêter prématurément.",{"type":32,"tag":33,"props":67,"children":68},{},[69,71,75,77,82],{"type":37,"value":70},"En A\u002FB fréquentiste, vous calculez la taille d'échantillon minimale pour atteindre une ",{"type":32,"tag":40,"props":72,"children":73},{},[74],{"type":37,"value":33},{"type":37,"value":76},"-value \u003C 0,05, puis attendez d'atteindre ce nombre. Or, la taille des cohortes varie quotidiennement dans les jeux mobiles : une nouvelle fonctionnalité augmente le DAU de 40 %, ou la saisonnalité estivale le réduit de 25 %. Le modèle bayésien interprète ces fluctuations comme des mises à jour du ",{"type":32,"tag":40,"props":78,"children":79},{},[80],{"type":37,"value":81},"prior",{"type":37,"value":83},", sans se laisser piéger par un plan de taille d'échantillon fixe.",{"type":32,"tag":33,"props":85,"children":86},{},[87],{"type":37,"value":88},"Exemple concret : dans un jeu avec 10 000 DAU, vous testez le prix du starter pack à $9.99. Le calcul fréquentiste requiert 42 000 utilisateurs pour détecter un lift de revenue de 5 % pendant 6 semaines. Le modèle bayésien, à la 3e semaine, affiche une moyenne posterior de $11.2 ARPPU pour le variant, $10.8 pour le contrôle, avec des intervalles de crédibilité qui ne se chevauchent pas — décision prise. Le test s'arrête. Les 3 semaines de revenue perdue sont récupérées.",{"type":32,"tag":90,"props":91,"children":93},"h3",{"id":92},"sélection-du-prior-et-segmentation",[94,96,101],{"type":37,"value":95},"Sélection du ",{"type":32,"tag":40,"props":97,"children":98},{},[99],{"type":37,"value":100},"Prior",{"type":37,"value":102}," et Segmentation",{"type":32,"tag":33,"props":104,"children":105},{},[106,108,112,114,118,120,124,126,130],{"type":37,"value":107},"Dans les tests bayésiens, le choix de la distribution ",{"type":32,"tag":40,"props":109,"children":110},{},[111],{"type":37,"value":81},{"type":37,"value":113}," n'est pas subjectif — il est fondé sur les données historiques. Si vous avez testé 8 paliers de prix entre $4.99 et $9.99 l'année précédente sur un jeu similaire, vous extrayez une distribution bêta ",{"type":32,"tag":40,"props":115,"children":116},{},[117],{"type":37,"value":81},{"type":37,"value":119}," de ces données. Le ",{"type":32,"tag":40,"props":121,"children":122},{},[123],{"type":37,"value":81},{"type":37,"value":125}," peut être faible (variance élevée) mais il est supérieur à un ",{"type":32,"tag":40,"props":127,"children":128},{},[129],{"type":37,"value":81},{"type":37,"value":131}," uniforme non informatif, car vous savez que le taux de conversion des whale ne descendra jamais sous 0,5 %.",{"type":32,"tag":33,"props":133,"children":134},{},[135,137,141,143,147,149,153],{"type":37,"value":136},"La segmentation renforce le ",{"type":32,"tag":40,"props":138,"children":139},{},[140],{"type":37,"value":81},{"type":37,"value":142}," : vous utilisez un ",{"type":32,"tag":40,"props":144,"children":145},{},[146],{"type":37,"value":81},{"type":37,"value":148}," non informatif pour les nouveaux utilisateurs, et un ",{"type":32,"tag":40,"props":150,"children":151},{},[152],{"type":37,"value":81},{"type":37,"value":154}," serré pour les utilisateurs avec 30+ jours de rétention. Un modèle bayésien hiérarchique estime simultanément les paramètres au niveau du segment et au niveau global — chaque segment utilise ses propres données tout en partageant la tendance globale. Cette approche prévient l'overfitting dans les petits segments.",{"type":32,"tag":48,"props":156,"children":158},{"id":157},"architecture-de-la-price-ladder-iap",[159],{"type":37,"value":160},"Architecture de la Price Ladder IAP",{"type":32,"tag":33,"props":162,"children":163},{},[164],{"type":37,"value":165},"Dans les jeux F2P, la price ladder n'est pas plate mais distribuée sur une échelle logarithmique : $0.99, $2.99, $4.99, $9.99, $19.99, $49.99, $99.99. Ces paliers ont une justification psychologique (charm pricing) mais surtout économique : chaque étape capture un segment de willingness-to-pay différent. Dans l'optimisation bayésienne, chaque palier possède sa propre posterior, et ils s'influencent mutuellement — si vous augmentez $4.99, la conversion à $2.99 peut baisser (downgrade), tandis que $9.99 augmente (upgrade).",{"type":32,"tag":33,"props":167,"children":168},{},[169,171,176],{"type":37,"value":170},"Dans le test d'une ladder complète, vous n'optimisez pas un prix unique mais toute l'escalade. Un algorithme ",{"type":32,"tag":40,"props":172,"children":173},{},[174],{"type":37,"value":175},"multi-armed bandit",{"type":37,"value":177}," traite chaque point de prix comme un « bras », utilisant Thompson Sampling pour puiser dans la posterior actuelle et sélectionner le prix d'expected revenue maximal. Les deux premières semaines, tous les bras reçoivent 14 % du trafic chacun (exploration uniforme). À partir de la 3e semaine, à mesure que la confiance posterior augmente, l'exploitation prend le pas.",{"type":32,"tag":33,"props":179,"children":180},{},[181],{"type":37,"value":182},"Exemple de scénario : ladder à 7 paliers, test de 21 jours. Jours 1-7, chaque prix reçoit 14 % du trafic. À partir du jour 8, le prix avec le plus haut (posterior mean × conversion rate) attire le trafic. Au jour 21, $4.99 reçoit 40 % du trafic, $9.99 reçoit 25 %, les autres 5-10 % chacun. En conclusion, les deux paliers restent actifs car ils génèrent tous deux une marginal revenue positive sans se cannabaliser.",{"type":32,"tag":90,"props":184,"children":186},{"id":185},"tarification-basée-sur-les-segments",[187],{"type":37,"value":188},"Tarification Basée sur les Segments",{"type":32,"tag":33,"props":190,"children":191},{},[192,194,199,201,206,208,212],{"type":37,"value":193},"Les segments whale\u002Fdolphin\u002Fminnow ne réagissent pas au même prix car l'élasticité est différente. Les whale (top 1 % des dépensiers) qui achètent des paquets à $99.99 voient leur conversion baisser de seulement 3 % si le prix augmente de 20 % — ",{"type":32,"tag":40,"props":195,"children":196},{},[197],{"type":37,"value":198},"inélastique",{"type":37,"value":200},". Les minnow (utilisateurs qui achètent $0.99 en premiers 7 jours) baissent leur conversion de 18 % pour une augmentation de 10 % du prix — ",{"type":32,"tag":40,"props":202,"children":203},{},[204],{"type":37,"value":205},"élastique",{"type":37,"value":207},". Le modèle bayésien encode cette élasticité dans le ",{"type":32,"tag":40,"props":209,"children":210},{},[211],{"type":37,"value":81},{"type":37,"value":213}," au niveau du segment.",{"type":32,"tag":33,"props":215,"children":216},{},[217,219,223],{"type":37,"value":218},"La segmentation utilise des features : jours depuis l'installation (D1\u002FD7\u002FD30), total dépensé, temps écoulé depuis le dernier IAP, fréquence des sessions, progression du niveau. Ces features construisent un ",{"type":32,"tag":40,"props":220,"children":221},{},[222],{"type":37,"value":81},{"type":37,"value":224}," de segment latent — un modèle hiérarchique estime également l'appartenance au segment. Ainsi, quand un nouvel utilisateur arrive, ses 24 premières heures de comportement permettent une prédiction de segment et un affichage de prix adapté.",{"type":32,"tag":33,"props":226,"children":227},{},[228,230,237],{"type":37,"value":229},"Dans le travail d'",{"type":32,"tag":231,"props":232,"children":234},"a",{"href":233},"\u002Ftr\u002Faso",[235],{"type":37,"value":236},"optimisation de l'App Store",{"type":37,"value":238}," chez Roibase, une segmentation similaire est utilisée : les résultats des tests créatifs varient selon le segment d'utilisateur — une même création affiche 8 % d'IPM sur iOS 16+ mais seulement 3 % sur iOS 15. L'intégration de l'ASO à l'optimisation IAP assure la cohérence de l'entonnoir — afficher le bon prix au bon utilisateur exige d'abord d'attirer le bon utilisateur.",{"type":32,"tag":48,"props":240,"children":242},{"id":241},"estimation-posterior-et-mécanisme-de-décision",[243],{"type":37,"value":244},"Estimation Posterior et Mécanisme de Décision",{"type":32,"tag":33,"props":246,"children":247},{},[248,250,255,257,261,263,267],{"type":37,"value":249},"Dans un test bayésien, la métrique de décision est la ",{"type":32,"tag":40,"props":251,"children":252},{},[253],{"type":37,"value":254},"probability of superiority",{"type":37,"value":256}," : P(treatment > control | data). Quand cette probabilité dépasse 95 %, le traitement gagne. La différence avec la ",{"type":32,"tag":40,"props":258,"children":259},{},[260],{"type":37,"value":33},{"type":37,"value":262},"-value fréquentiste est fondamentale : une ",{"type":32,"tag":40,"props":264,"children":265},{},[266],{"type":37,"value":33},{"type":37,"value":268},"-value mesure l'extrêmité des données sous l'hypothèse nulle, tandis que la probabilité posterior estime directement « la probabilité que le treatment soit meilleur ».",{"type":32,"tag":33,"props":270,"children":271},{},[272,274,278],{"type":37,"value":273},"Pour calculer la posterior, si vous utilisez un ",{"type":32,"tag":40,"props":275,"children":276},{},[277],{"type":37,"value":81},{"type":37,"value":279}," conjugué, il existe une solution analytique (bêta-binomiale). Sinon, une simulation MCMC (Markov Chain Monte Carlo) est nécessaire. Dans les tests gaming mobile, une combinaison binomiale pour la conversion + distribution lognormale pour la revenue fonctionne bien. PyMC3 ou Stan exécutent 10 000 itérations MCMC en 30 secondes ; la mise à jour quotidienne des données rafraîchit la posterior.",{"type":32,"tag":33,"props":281,"children":282},{},[283,285,290],{"type":37,"value":284},"Le seuil de décision peut être fixé à 90 % au lieu de 95 % — en phase de croissance agressive, 90 % suffit ; dans un jeu mature, 95 % est préférable. Le seuil bas augmente le risque de faux positif mais raccourcit le test. L'",{"type":32,"tag":40,"props":286,"children":287},{},[288],{"type":37,"value":289},"Expected Value of Information",{"type":37,"value":291}," (EVI) calcule le seuil optimal : on confronte le coût d'une semaine supplémentaire de test au coût d'une décision erronée, et on trouve l'équilibre.",{"type":32,"tag":90,"props":293,"children":295},{"id":294},"structure-du-test-bayésien-multi-variant",[296],{"type":37,"value":297},"Structure du Test Bayésien Multi-Variant",{"type":32,"tag":33,"props":299,"children":300},{},[301],{"type":37,"value":302},"Un test de prix IAP comprend souvent 3+ variantes : contrôle ($4.99), traitement A ($5.99), traitement B ($6.99). En test A\u002FB fréquentiste, le problème des comparaisons multiples surgit ; la correction de Bonferroni multiplie la taille d'échantillon. En bayésien, chaque variante a sa propre posterior, les comparaisons par paires se font simultanément. Au lieu de sélectionner la variante avec la plus grande moyenne posterior, vous maximisez la revenue attendue : (probabilité de gagner) × (revenue attendue) pour chaque variante.",{"type":32,"tag":33,"props":304,"children":305},{},[306],{"type":37,"value":307},"La stratégie Thompson Sampling fonctionne ainsi : chaque jour, tirez un échantillon de la posterior de chaque variante, sélectionnez l'échantillon le plus élevé et envoyez le trafic vers cette variante. Cette stratégie équilibre automatiquement exploration et exploitation — la distribution du trafic est quasi-uniforme quand l'incertitude posterior est haute (premiers jours), puis bascule vers la variante gagnante.",{"type":32,"tag":33,"props":309,"children":310},{},[311],{"type":37,"value":312},"Snippet de code (modèle bêta-binomial avec PyMC3) :",{"type":32,"tag":314,"props":315,"children":319},"pre",{"className":316,"code":317,"language":318,"meta":16,"style":16},"language-python shiki shiki-themes github-dark","import pymc3 as pm\n\nwith pm.Model() as iap_model:\n    # Prior : bêta uniforme\n    p_control = pm.Beta('p_control', alpha=1, beta=1)\n    p_treatment = pm.Beta('p_treatment', alpha=1, beta=1)\n    \n    # Likelihood\n    obs_control = pm.Binomial('obs_control', n=n_control, p=p_control, observed=conversions_control)\n    obs_treatment = pm.Binomial('obs_treatment', n=n_treatment, p=p_treatment, observed=conversions_treatment)\n    \n    # Posterior sampling\n    trace = pm.sample(10000, return_inferencedata=False)\n    \n    # Probability of superiority\n    prob_superiority = (trace['p_treatment'] > trace['p_control']).mean()\n","python",[320],{"type":32,"tag":321,"props":322,"children":323},"code",{"__ignoreMap":16},[324,352,362,385,395,463,521,530,539,606,671,679,688,733,741,750],{"type":32,"tag":325,"props":326,"children":329},"span",{"class":327,"line":328},"line",1,[330,336,342,347],{"type":32,"tag":325,"props":331,"children":333},{"style":332},"--shiki-default:#F97583",[334],{"type":37,"value":335},"import",{"type":32,"tag":325,"props":337,"children":339},{"style":338},"--shiki-default:#E1E4E8",[340],{"type":37,"value":341}," pymc3 ",{"type":32,"tag":325,"props":343,"children":344},{"style":332},[345],{"type":37,"value":346},"as",{"type":32,"tag":325,"props":348,"children":349},{"style":338},[350],{"type":37,"value":351}," pm\n",{"type":32,"tag":325,"props":353,"children":355},{"class":327,"line":354},2,[356],{"type":32,"tag":325,"props":357,"children":359},{"emptyLinePlaceholder":358},true,[360],{"type":37,"value":361},"\n",{"type":32,"tag":325,"props":363,"children":365},{"class":327,"line":364},3,[366,371,376,380],{"type":32,"tag":325,"props":367,"children":368},{"style":332},[369],{"type":37,"value":370},"with",{"type":32,"tag":325,"props":372,"children":373},{"style":338},[374],{"type":37,"value":375}," pm.Model() ",{"type":32,"tag":325,"props":377,"children":378},{"style":332},[379],{"type":37,"value":346},{"type":32,"tag":325,"props":381,"children":382},{"style":338},[383],{"type":37,"value":384}," iap_model:\n",{"type":32,"tag":325,"props":386,"children":388},{"class":327,"line":387},4,[389],{"type":32,"tag":325,"props":390,"children":392},{"style":391},"--shiki-default:#6A737D",[393],{"type":37,"value":394},"    # Prior : bêta uniforme\n",{"type":32,"tag":325,"props":396,"children":398},{"class":327,"line":397},5,[399,404,409,414,420,425,431,435,441,445,450,454,458],{"type":32,"tag":325,"props":400,"children":401},{"style":338},[402],{"type":37,"value":403},"    p_control ",{"type":32,"tag":325,"props":405,"children":406},{"style":332},[407],{"type":37,"value":408},"=",{"type":32,"tag":325,"props":410,"children":411},{"style":338},[412],{"type":37,"value":413}," pm.Beta(",{"type":32,"tag":325,"props":415,"children":417},{"style":416},"--shiki-default:#9ECBFF",[418],{"type":37,"value":419},"'p_control'",{"type":32,"tag":325,"props":421,"children":422},{"style":338},[423],{"type":37,"value":424},", ",{"type":32,"tag":325,"props":426,"children":428},{"style":427},"--shiki-default:#FFAB70",[429],{"type":37,"value":430},"alpha",{"type":32,"tag":325,"props":432,"children":433},{"style":332},[434],{"type":37,"value":408},{"type":32,"tag":325,"props":436,"children":438},{"style":437},"--shiki-default:#79B8FF",[439],{"type":37,"value":440},"1",{"type":32,"tag":325,"props":442,"children":443},{"style":338},[444],{"type":37,"value":424},{"type":32,"tag":325,"props":446,"children":447},{"style":427},[448],{"type":37,"value":449},"beta",{"type":32,"tag":325,"props":451,"children":452},{"style":332},[453],{"type":37,"value":408},{"type":32,"tag":325,"props":455,"children":456},{"style":437},[457],{"type":37,"value":440},{"type":32,"tag":325,"props":459,"children":460},{"style":338},[461],{"type":37,"value":462},")\n",{"type":32,"tag":325,"props":464,"children":466},{"class":327,"line":465},6,[467,472,476,480,485,489,493,497,501,505,509,513,517],{"type":32,"tag":325,"props":468,"children":469},{"style":338},[470],{"type":37,"value":471},"    p_treatment ",{"type":32,"tag":325,"props":473,"children":474},{"style":332},[475],{"type":37,"value":408},{"type":32,"tag":325,"props":477,"children":478},{"style":338},[479],{"type":37,"value":413},{"type":32,"tag":325,"props":481,"children":482},{"style":416},[483],{"type":37,"value":484},"'p_treatment'",{"type":32,"tag":325,"props":486,"children":487},{"style":338},[488],{"type":37,"value":424},{"type":32,"tag":325,"props":490,"children":491},{"style":427},[492],{"type":37,"value":430},{"type":32,"tag":325,"props":494,"children":495},{"style":332},[496],{"type":37,"value":408},{"type":32,"tag":325,"props":498,"children":499},{"style":437},[500],{"type":37,"value":440},{"type":32,"tag":325,"props":502,"children":503},{"style":338},[504],{"type":37,"value":424},{"type":32,"tag":325,"props":506,"children":507},{"style":427},[508],{"type":37,"value":449},{"type":32,"tag":325,"props":510,"children":511},{"style":332},[512],{"type":37,"value":408},{"type":32,"tag":325,"props":514,"children":515},{"style":437},[516],{"type":37,"value":440},{"type":32,"tag":325,"props":518,"children":519},{"style":338},[520],{"type":37,"value":462},{"type":32,"tag":325,"props":522,"children":524},{"class":327,"line":523},7,[525],{"type":32,"tag":325,"props":526,"children":527},{"style":338},[528],{"type":37,"value":529},"    \n",{"type":32,"tag":325,"props":531,"children":533},{"class":327,"line":532},8,[534],{"type":32,"tag":325,"props":535,"children":536},{"style":391},[537],{"type":37,"value":538},"    # Likelihood\n",{"type":32,"tag":325,"props":540,"children":541},{"class":327,"line":26},[542,547,551,556,561,565,570,574,579,583,587,592,597,601],{"type":32,"tag":325,"props":543,"children":544},{"style":338},[545],{"type":37,"value":546},"    obs_control ",{"type":32,"tag":325,"props":548,"children":549},{"style":332},[550],{"type":37,"value":408},{"type":32,"tag":325,"props":552,"children":553},{"style":338},[554],{"type":37,"value":555}," pm.Binomial(",{"type":32,"tag":325,"props":557,"children":558},{"style":416},[559],{"type":37,"value":560},"'obs_control'",{"type":32,"tag":325,"props":562,"children":563},{"style":338},[564],{"type":37,"value":424},{"type":32,"tag":325,"props":566,"children":567},{"style":427},[568],{"type":37,"value":569},"n",{"type":32,"tag":325,"props":571,"children":572},{"style":332},[573],{"type":37,"value":408},{"type":32,"tag":325,"props":575,"children":576},{"style":338},[577],{"type":37,"value":578},"n_control, ",{"type":32,"tag":325,"props":580,"children":581},{"style":427},[582],{"type":37,"value":33},{"type":32,"tag":325,"props":584,"children":585},{"style":332},[586],{"type":37,"value":408},{"type":32,"tag":325,"props":588,"children":589},{"style":338},[590],{"type":37,"value":591},"p_control, ",{"type":32,"tag":325,"props":593,"children":594},{"style":427},[595],{"type":37,"value":596},"observed",{"type":32,"tag":325,"props":598,"children":599},{"style":332},[600],{"type":37,"value":408},{"type":32,"tag":325,"props":602,"children":603},{"style":338},[604],{"type":37,"value":605},"conversions_control)\n",{"type":32,"tag":325,"props":607,"children":609},{"class":327,"line":608},10,[610,615,619,623,628,632,636,640,645,649,653,658,662,666],{"type":32,"tag":325,"props":611,"children":612},{"style":338},[613],{"type":37,"value":614},"    obs_treatment ",{"type":32,"tag":325,"props":616,"children":617},{"style":332},[618],{"type":37,"value":408},{"type":32,"tag":325,"props":620,"children":621},{"style":338},[622],{"type":37,"value":555},{"type":32,"tag":325,"props":624,"children":625},{"style":416},[626],{"type":37,"value":627},"'obs_treatment'",{"type":32,"tag":325,"props":629,"children":630},{"style":338},[631],{"type":37,"value":424},{"type":32,"tag":325,"props":633,"children":634},{"style":427},[635],{"type":37,"value":569},{"type":32,"tag":325,"props":637,"children":638},{"style":332},[639],{"type":37,"value":408},{"type":32,"tag":325,"props":641,"children":642},{"style":338},[643],{"type":37,"value":644},"n_treatment, ",{"type":32,"tag":325,"props":646,"children":647},{"style":427},[648],{"type":37,"value":33},{"type":32,"tag":325,"props":650,"children":651},{"style":332},[652],{"type":37,"value":408},{"type":32,"tag":325,"props":654,"children":655},{"style":338},[656],{"type":37,"value":657},"p_treatment, ",{"type":32,"tag":325,"props":659,"children":660},{"style":427},[661],{"type":37,"value":596},{"type":32,"tag":325,"props":663,"children":664},{"style":332},[665],{"type":37,"value":408},{"type":32,"tag":325,"props":667,"children":668},{"style":338},[669],{"type":37,"value":670},"conversions_treatment)\n",{"type":32,"tag":325,"props":672,"children":674},{"class":327,"line":673},11,[675],{"type":32,"tag":325,"props":676,"children":677},{"style":338},[678],{"type":37,"value":529},{"type":32,"tag":325,"props":680,"children":682},{"class":327,"line":681},12,[683],{"type":32,"tag":325,"props":684,"children":685},{"style":391},[686],{"type":37,"value":687},"    # Posterior sampling\n",{"type":32,"tag":325,"props":689,"children":691},{"class":327,"line":690},13,[692,697,701,706,711,715,720,724,729],{"type":32,"tag":325,"props":693,"children":694},{"style":338},[695],{"type":37,"value":696},"    trace ",{"type":32,"tag":325,"props":698,"children":699},{"style":332},[700],{"type":37,"value":408},{"type":32,"tag":325,"props":702,"children":703},{"style":338},[704],{"type":37,"value":705}," pm.sample(",{"type":32,"tag":325,"props":707,"children":708},{"style":437},[709],{"type":37,"value":710},"10000",{"type":32,"tag":325,"props":712,"children":713},{"style":338},[714],{"type":37,"value":424},{"type":32,"tag":325,"props":716,"children":717},{"style":427},[718],{"type":37,"value":719},"return_inferencedata",{"type":32,"tag":325,"props":721,"children":722},{"style":332},[723],{"type":37,"value":408},{"type":32,"tag":325,"props":725,"children":726},{"style":437},[727],{"type":37,"value":728},"False",{"type":32,"tag":325,"props":730,"children":731},{"style":338},[732],{"type":37,"value":462},{"type":32,"tag":325,"props":734,"children":736},{"class":327,"line":735},14,[737],{"type":32,"tag":325,"props":738,"children":739},{"style":338},[740],{"type":37,"value":529},{"type":32,"tag":325,"props":742,"children":744},{"class":327,"line":743},15,[745],{"type":32,"tag":325,"props":746,"children":747},{"style":391},[748],{"type":37,"value":749},"    # Probability of superiority\n",{"type":32,"tag":325,"props":751,"children":753},{"class":327,"line":752},16,[754,759,763,768,772,777,782,787,791],{"type":32,"tag":325,"props":755,"children":756},{"style":338},[757],{"type":37,"value":758},"    prob_superiority ",{"type":32,"tag":325,"props":760,"children":761},{"style":332},[762],{"type":37,"value":408},{"type":32,"tag":325,"props":764,"children":765},{"style":338},[766],{"type":37,"value":767}," (trace[",{"type":32,"tag":325,"props":769,"children":770},{"style":416},[771],{"type":37,"value":484},{"type":32,"tag":325,"props":773,"children":774},{"style":338},[775],{"type":37,"value":776},"] ",{"type":32,"tag":325,"props":778,"children":779},{"style":332},[780],{"type":37,"value":781},">",{"type":32,"tag":325,"props":783,"children":784},{"style":338},[785],{"type":37,"value":786}," trace[",{"type":32,"tag":325,"props":788,"children":789},{"style":416},[790],{"type":37,"value":419},{"type":32,"tag":325,"props":792,"children":793},{"style":338},[794],{"type":37,"value":795},"]).mean()\n",{"type":32,"tag":33,"props":797,"children":798},{},[799,801,805,807,813],{"type":37,"value":800},"Ce modèle optimise le taux de conversion. Pour optimiser la revenue, ajoutez un ",{"type":32,"tag":40,"props":802,"children":803},{},[804],{"type":37,"value":81},{"type":37,"value":806}," lognormal et calculez la posterior jointe de ",{"type":32,"tag":321,"props":808,"children":810},{"className":809},[],[811],{"type":37,"value":812},"p × revenue_mean",{"type":37,"value":814},".",{"type":32,"tag":48,"props":816,"children":818},{"id":817},"migration-de-segment-et-impact-long-terme",[819],{"type":37,"value":820},"Migration de Segment et Impact Long-Terme",{"type":32,"tag":33,"props":822,"children":823},{},[824,826,830,832,836,838,842],{"type":37,"value":825},"L'optimisation des prix n'est pas un test unique mais un processus continu. Les utilisateurs changent de segment : un minnow aujourd'hui peut être dolphin dans 30 jours. Le modèle bayésien ne capture pas cette migration car il utilise un ",{"type":32,"tag":40,"props":827,"children":828},{},[829],{"type":37,"value":81},{"type":37,"value":831}," statique. La solution : mise à jour dynamique du ",{"type":32,"tag":40,"props":833,"children":834},{},[835],{"type":37,"value":81},{"type":37,"value":837}," — tous les 30 jours, la posterior devient un nouveau ",{"type":32,"tag":40,"props":839,"children":840},{},[841],{"type":37,"value":81},{"type":37,"value":843}," combiné aux nouvelles données.",{"type":32,"tag":33,"props":845,"children":846},{},[847],{"type":37,"value":848},"Pour mesurer l'impact long-terme, la courbe de rétention des cohortes est modélisée avec l'analyse bayésienne de survie. Si une hausse de prix abaisse la rétention D7 de 2 % mais élève le LTV de $12 à $14, c'est un gain net. Le modèle de survie utilise la distribution de Weibull pour estimer les paramètres de forme et d'échelle ; les vérifications prédictives posterior fournissent une prévision du LTV à 90 jours.",{"type":32,"tag":33,"props":850,"children":851},{},[852,854,858,860,865,867,872],{"type":37,"value":853},"Le test d'impact sur la rétention prend 6-8 semaines car il faut attendre le signal de rétention D30. L'approche bayésienne prédit D30 à partir de données D7 — le ",{"type":32,"tag":40,"props":855,"children":856},{},[857],{"type":37,"value":81},{"type":37,"value":859}," est fondé sur les taux de transition D7→D30 des cohortes passées. Cela permet de détecter un signal précoce dès la 3e semaine : si la moyenne posterior pour D30 est 18 % avec un IC 95 % de ",{"type":32,"tag":325,"props":861,"children":862},{},[863],{"type":37,"value":864},"16 %, 20 %",{"type":37,"value":866},", le test continue ; si ",{"type":32,"tag":325,"props":868,"children":869},{},[870],{"type":37,"value":871},"14 %, 16 %",{"type":37,"value":873},", la hausse de prix casse la rétention, et on revient au prix précédent.",{"type":32,"tag":48,"props":875,"children":877},{"id":876},"économie-du-jeu-et-dynamiques-plateformes",[878],{"type":37,"value":879},"Économie du Jeu et Dynamiques Plateformes",{"type":32,"tag":33,"props":881,"children":882},{},[883,885,889],{"type":37,"value":884},"Les utilisateurs iOS et Android réagissent différemment à la même price ladder. Les utilisateurs iOS affichent en moyenne 23 % d'ARPPU plus élevé ; le même prix $4.99 produit 3,2 % de conversion sur iOS mais 2,1 % sur Android. Le modèle bayésien intègre la plateforme comme un facteur hiérarchique — chaque plateforme a son propre ",{"type":32,"tag":40,"props":886,"children":887},{},[888],{"type":37,"value":81},{"type":37,"value":890}," de segment mais partage la tendance mondiale.",{"type":32,"tag":33,"props":892,"children":893},{},[894,896,900],{"type":37,"value":895},"Le système de tarification de l'App Store d'Apple (Tier 1 = $0.99, Tier 5 = $4.99...) limite la flexibilité. Entre les paliers, au lieu de tester la posterior, on effectue une grille de recherche : entre Tier 3\u002F4\u002F5, lequel maximise la posterior revenue attendue ? Google Play est plus flexible (tarification arbitraire) mais le taux de conversion Android est plus volatil — les tests Android conservent un ",{"type":32,"tag":40,"props":897,"children":898},{},[899],{"type":37,"value":81},{"type":37,"value":901}," 30 % plus large que iOS.",{"type":32,"tag":33,"props":903,"children":904},{},[905,907,911],{"type":37,"value":906},"Les fluctuations monétaires affectent aussi la posterior : quand le cours TRY passe de ₺25 à ₺35 pour 1 USD, le prix ₺49.99 descend de $2 à $1.43 en terms réels. Le modèle utilise la revenue ajustée par devise, avec la posterior calculée en USD. Sur les marchés émergents, une tarification ajustée par PPA exige des ",{"type":32,"tag":40,"props":908,"children":909},{},[910],{"type":37,"value":44},{"type":37,"value":912}," distincts — le même jeu affiche $4.99 aux États-Unis et R$9.90 au Brésil (équivalent PPA ~$1.80).",{"type":32,"tag":33,"props":914,"children":915},{},[916,918,924],{"type":37,"value":917},"Dans le ",{"type":32,"tag":231,"props":919,"children":921},{"href":920},"\u002Ftr\u002Fpremiumyayinci",[922],{"type":37,"value":923},"Programme Éditeur Premium",{"type":37,"value":925},", les campagnes UA alimentent aussi les résultats du test de prix : pour un segment à LTV élevé, on augmente l'enchère CPM ; pour un segment à faible conversion, on la réduit. Quand le modèle bayésien IAP est intégré à la stratégie d'enchères UA, l'optimisation du ROI au niveau du portefeuille devient possible —",{"type":32,"tag":927,"props":928,"children":929},"style",{},[930],{"type":37,"value":931},"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":364,"depth":364,"links":933},[934,938,941,944,945],{"id":50,"depth":354,"text":53,"children":935},[936],{"id":92,"depth":364,"text":937},"Sélection du Prior et Segmentation",{"id":157,"depth":354,"text":160,"children":939},[940],{"id":185,"depth":364,"text":188},{"id":241,"depth":354,"text":244,"children":942},[943],{"id":294,"depth":364,"text":297},{"id":817,"depth":354,"text":820},{"id":876,"depth":354,"text":879},"markdown","content:fr:gaming:optimisation-bayesienne-des-prix-f2p-mobile.md","content","fr\u002Fgaming\u002Foptimisation-bayesienne-des-prix-f2p-mobile.md","fr\u002Fgaming\u002Foptimisation-bayesienne-des-prix-f2p-mobile","md",1785967483054]