{"id":2735,"date":"2026-08-20T12:41:48","date_gmt":"2026-08-20T10:41:48","guid":{"rendered":"https:\/\/mairateam.com\/?post_type=glossary&p=2735"},"modified":"2026-08-20T15:49:18","modified_gmt":"2026-08-20T13:49:18","slug":"marketing-mix-modeling","status":"publish","type":"glossary","link":"https:\/\/mairateam.com\/cs\/slovnik\/marketing-mix-modeling\/","title":{"rendered":"Marketing mix modeling"},"content":{"rendered":"

Marketing mix modeling<\/strong>, zkr\u00e1cen\u011b MMM, je statistick\u00fd model, kter\u00fd z historick\u00fdch \u010dasov\u00fdch \u0159ad odhadne, jak jednotliv\u00e9 kan\u00e1ly, sez\u00f3na, ceny a dal\u0161\u00ed vlivy p\u0159isp\u00edvaj\u00ed k tr\u017eb\u00e1m. Nepracuje s cookies ani se sledov\u00e1n\u00edm jednotlivc\u016f, pot\u0159ebuje jen agregovan\u00e1 data: v\u00fddaje po kan\u00e1lech, tr\u017eby, sez\u00f3nnost. Pr\u00e1v\u011b proto se po \u00fapadku cookie m\u011b\u0159en\u00ed vrac\u00ed jako hlavn\u00ed n\u00e1stroj pro strategick\u00e9 rozd\u011blen\u00ed rozpo\u010dtu.<\/p>\n

V kostce<\/h2>\n\n\n\n\n\n\n
Co to je<\/strong><\/td>\nStatistick\u00fd odhad p\u0159\u00edsp\u011bvku kan\u00e1l\u016f k tr\u017eb\u00e1m z historick\u00fdch dat<\/td>\n<\/tr>\n
Co pot\u0159ebuje<\/strong><\/td>\nAgregovan\u00e9 \u010dasov\u00e9 \u0159ady: v\u00fddaje, tr\u017eby, sez\u00f3nnost, promoakce, ide\u00e1ln\u011b 2 a\u017e 3 roky<\/td>\n<\/tr>\n
Co nepot\u0159ebuje<\/strong><\/td>\nCookies, pixely, souhlasy, sledov\u00e1n\u00ed jednotlivc\u016f<\/td>\n<\/tr>\n
K \u010demu slou\u017e\u00ed<\/strong><\/td>\nStrategick\u00e9 rozhodnut\u00ed o rozd\u011blen\u00ed rozpo\u010dtu mezi kan\u00e1ly, ne denn\u00ed optimalizace<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Jak to funguje<\/h2>\n

Vstupy a rovnice modelu<\/h3>\n

Z\u00e1kladem je regrese na \u010dasov\u00fdch \u0159ad\u00e1ch. Vysv\u011btlovan\u00e1 prom\u011bnn\u00e1 jsou tr\u017eby (nebo objedn\u00e1vky) po t\u00fddnech, vysv\u011btluj\u00edc\u00ed prom\u011bnn\u00e9 jsou v\u00fddaje nebo imprese ka\u017ed\u00e9ho kan\u00e1lu a k nim kontroln\u00ed prom\u011bnn\u00e9: sez\u00f3nnost, sv\u00e1tky, promoakce, zm\u011bny cen, po\u010das\u00ed, aktivita konkurence, p\u0159\u00edpadn\u011b makroekonomick\u00e9 ukazatele. Model hled\u00e1 koeficienty, se kter\u00fdmi sou\u010det p\u0159\u00edsp\u011bvk\u016f nejl\u00e9pe vysv\u011btl\u00ed skute\u010dn\u00e9 tr\u017eby. \u010c\u00e1st tr\u017eeb, kterou nevysv\u011btl\u00ed \u017e\u00e1dn\u00fd kan\u00e1l, je z\u00e1kladna (baseline): organick\u00e1 popt\u00e1vka, s\u00edla zna\u010dky, opakovan\u00e9 n\u00e1kupy. Zjednodu\u0161en\u011b: tr\u017eby v t\u00fddnu t = z\u00e1kladna + p\u0159\u00edsp\u011bvek kan\u00e1lu 1 + p\u0159\u00edsp\u011bvek kan\u00e1lu 2 + ... + vliv kontroln\u00edch prom\u011bnn\u00fdch + \u0161um.<\/p>\n

Adstock a saturace<\/h3>\n

Dv\u011b transformace odli\u0161uj\u00ed MMM od oby\u010dejn\u00e9 regrese. Adstock<\/strong> zachycuje dozn\u00edv\u00e1n\u00ed: reklama z jednoho t\u00fddne p\u016fsob\u00ed i v t\u00fddnech n\u00e1sleduj\u00edc\u00edch, proto se v\u00fddaje p\u0159ed vstupem do modelu rozlo\u017e\u00ed v \u010dase. Modelov\u00fd p\u0159\u00edklad: p\u0159i koeficientu p\u0159enosu 0,5 se z v\u00fddaje 100 000 K\u010d v prvn\u00edm t\u00fddnu zapo\u010d\u00edt\u00e1 100 000 K\u010d, do druh\u00e9ho t\u00fddne se p\u0159enese 50 000 K\u010d, do t\u0159et\u00edho 25 000 K\u010d a d\u00e1l kles\u00e1. Kan\u00e1ly s dlouh\u00fdm dozn\u00edv\u00e1n\u00edm (TV, video, brand) m\u00edvaj\u00ed vy\u0161\u0161\u00ed koeficient ne\u017e search. Saturace<\/strong> zachycuje klesaj\u00edc\u00ed v\u00fdnos: k\u0159ivka odezvy (typicky S-k\u0159ivka nebo Hillova funkce) \u0159\u00edk\u00e1, \u017ee prvn\u00ed koruny v kan\u00e1lu p\u0159in\u00e1\u0161ej\u00ed v\u00edc ne\u017e ty posledn\u00ed. Model tedy neodhaduje jeden pr\u016fm\u011brn\u00fd ROAS kan\u00e1lu, ale celou k\u0159ivku, ze kter\u00e9 se \u010dte mezn\u00ed v\u00fdnos: kolik p\u0159inese dal\u0161\u00ed koruna p\u0159i aktu\u00e1ln\u00ed \u00farovni v\u00fddaj\u016f.<\/p>\n

Kalibrace a v\u00fdstupy<\/h3>\n

Regrese sama o sob\u011b nerozli\u0161\u00ed p\u0159\u00ed\u010dinu od korelace: kan\u00e1ly, jejich\u017e v\u00fddaje rostou spole\u010dn\u011b se sez\u00f3nou, model od sebe neum\u00ed odli\u0161it. Proto se model kalibruje: koeficient\u016fm se zadaj\u00ed apriorn\u00ed meze (priory) z geo lift test\u016f a holdout test\u016f, nap\u0159\u00edklad \u017ee inkrement\u00e1ln\u00ed ROAS kan\u00e1lu vy\u0161el mezi 1,5 a 2,5. Model se d\u00e1le ov\u011b\u0159uje na odlo\u017een\u00fdch t\u00fddnech (out of sample), kde nesm\u00ed zn\u00e1t v\u00fdsledek. V\u00fdstupem je rozklad tr\u017eeb podle kan\u00e1l\u016f a z\u00e1kladny, k\u0159ivka odezvy pro ka\u017ed\u00fd kan\u00e1l, mezn\u00ed ROAS a z n\u011bj odvozen\u00fd n\u00e1vrh p\u0159erozd\u011blen\u00ed rozpo\u010dtu p\u0159i stejn\u00e9m celkov\u00e9m v\u00fddaji nebo pro c\u00edlov\u00fd objem tr\u017eeb. Odhady maj\u00ed interval nejistoty a ten pat\u0159\u00ed do ka\u017ed\u00e9 prezentace v\u00fdsledk\u016f, jinak se z odhadu stane fale\u0161n\u00e1 p\u0159esnost.<\/p>\n

Kdy MMM d\u00e1v\u00e1 smysl a kdy ne<\/h2>\n

MMM odpov\u00edd\u00e1 na ot\u00e1zku, kterou atribuce z pixel\u016f zodpov\u011bd\u011bt neum\u00ed: co by se stalo s tr\u017ebami, kdybychom p\u0159esunuli rozpo\u010det mezi kan\u00e1ly, v\u010detn\u011b t\u011bch, kter\u00e9 pixel nevid\u00ed, jako je televize, r\u00e1dio nebo brandov\u00e1 kampa\u0148. Da\u0148 za to je hrubost: model pracuje po t\u00fddnech a kan\u00e1lech, ne po kampan\u00edch a sestav\u00e1ch. Pot\u0159ebuje tak\u00e9 dostate\u010dn\u00fd rozpo\u010det a variabilitu dat: kdy\u017e se v\u00fddaje nikdy nem\u011bnily, model nem\u00e1 z \u010deho odhadovat. Pro men\u0161\u00ed \u00fa\u010dty s jedn\u00edm a\u017e dv\u011bma kan\u00e1ly je MMM zbyte\u010dn\u011b t\u011b\u017ek\u00fd n\u00e1stroj a lep\u0161\u00ed slu\u017ebu ud\u011blaj\u00ed inkrement\u00e1ln\u00ed testy.<\/p>\n

Z na\u0161\u00ed praxe<\/h2>\n

Pro m\u011b\u0159en\u00ed skute\u010dn\u00e9ho p\u0159\u00ednosu kan\u00e1l\u016f pou\u017e\u00edv\u00e1me vlastn\u00ed platformu DiagnostIQ, kter\u00e1 stoj\u00ed na stejn\u00e9m principu jako MMM: na inkrement\u00e1ln\u00edm dopadu m\u00edsto p\u0159ipisov\u00e1n\u00ed konverz\u00ed. Konkr\u00e9tn\u00ed p\u0159\u00edklad v\u00fdsledku: inkrement\u00e1ln\u00ed test TikToku u e-commerce klienta uk\u00e1zal 15 procent tr\u017eeb nav\u00edc, kter\u00e9 by bez kan\u00e1lu nevznikly, p\u0159esto\u017ee atribu\u010dn\u00ed pohled z pixel\u016f kan\u00e1l podce\u0148oval. P\u0159esn\u011b tento typ rozporu je d\u016fvod, pro\u010d se strategick\u00e9 rozpo\u010dty nemaj\u00ed \u0159\u00eddit jen atribuc\u00ed.<\/p>\n

\u010cast\u00e9 chyby<\/h2>\n