加入新推出的
Discord 社区,展开实时讨论,获得同行支持,并直接与 Meridian 团队互动!
更新模型
使用集合让一切井井有条
根据您的偏好保存内容并对其进行分类。
刷新频率
您可以根据需要随时刷新模型。模型选择和调优通常是一个持续迭代的过程,可能需要随着新数据的出现而刷新。您可以考虑每季度、每年或以与营销预算决策制定流程一致的频率更新模型。
建议:将新数据添加到旧数据中,以整合新数据
我们建议您将新数据添加到旧数据中,并运行 Meridian。
您应该考虑是否要舍弃最旧的数据,以容纳新数据。为了使数据保持在营销组合建模分析 (MMM) 通常采用的 2-3 年的时间窗口内,这可能是必要的。在模型中,Meridian 并未将媒体效果设为随时间而变化。因此,决定在附加新数据时丢弃旧数据需要您在方差与偏差之间进行权衡。时间窗口越长,方差就越小,因为您有了更多的数据;但是,如果媒体效果和策略随着时间的推移发生了巨大变化,这可能会增加偏差。
请注意,MMM 估计值通常方差高。这意味着,即使是整合相对较少的新数据,也可能会对模型的结果产生显著影响。因此,在新模型中设置先验,以使新模型的后验与旧模型的后验相匹配,可能有合理的业务原因。我们建议您根据先验知识和直觉来设置先验,而这种直觉可以参考以往的 MMM 结果。至于您希望以往的 MMM 结果在多大程度上影响您的先验知识和直觉,则由您自己决定。不过,要考虑到,设置与以往的 MMM 结果相匹配的先验,实际上是对旧数据进行了两次计算。
替代方案:将新数据独立地纳入模型并使用先验
有些人可能会考虑仅基于新数据拟合模型来整合新数据,而不混入旧模型中所用的数据。虽然从技术上讲这是可行的,但即使是少量数据(例如一个季度的数据),通常也不建议这样做。
如果将新数据纳入模型,而完全不混入旧数据,将无法妥善考虑滞后效应。Meridian 允许媒体数据包含比 KPI 数据和控制变量数据更多的(更早的)时间段。这样,就可以从 KPI 数据的第一个时间段开始,更准确地对滞后效应进行模型分析。在可能的情况下,最好包含 KPI 数据的第一个时间段之前的 max_lag 个媒体数据时间段。
如果新数据较少,可能不足以让模型得出结论(请参阅“所需数据量”)。您可能希望通过使用参考了旧模型后验的先验分布,来整合旧数据中的信息。虽然所有形参的完整联合后验分布在理论上包含旧数据中的所有信息(将其用作新数据的先验与拟合同时包含新旧数据的新模型类似),但 Meridian 会对各个形参使用独立的先验分布。因此,即使各个形参的后验分布作为其先验传递,也可能无法完全捕获完整的联合后验分布,后者考虑了形参之间的相互依赖关系。此外,贝叶斯模型要求为每个形参提供形参先验分布。MCMC 采样可提供来自后验的经验样本,该样本可能具有或不具有适合直接用作先验的形参近似值。
如未另行说明,那么本页面中的内容已根据知识共享署名 4.0 许可获得了许可,并且代码示例已根据 Apache 2.0 许可获得了许可。有关详情,请参阅 Google 开发者网站政策。Java 是 Oracle 和/或其关联公司的注册商标。
最后更新时间 (UTC):2025-08-04。
[null,null,["最后更新时间 (UTC):2025-08-04。"],[[["\u003cp\u003eModel refresh frequency should align with data frequency and marketing team's decision-making timeframe (e.g., quarterly).\u003c/p\u003e\n"],["\u003cp\u003eExpanding the data window with each refresh allows older data to influence newer estimates while balancing bias and variance.\u003c/p\u003e\n"],["\u003cp\u003eEven small data additions can significantly impact results due to the inherent high-variance nature of MMM estimates.\u003c/p\u003e\n"],["\u003cp\u003ePrior settings can be adjusted to balance new and old data influences, informed by past results and business intuition.\u003c/p\u003e\n"]]],["Model refreshing frequency should align with data frequency and marketing decision timelines, such as quarterly updates for quarterly decisions. Appending new data reduces variance but may introduce bias if media strategies change. Appending small amounts of data can significantly impact estimates due to their high variance. When appending, setting priors to match previous results can align old and new data, although this risks double-counting data. Prior knowledge should influence prior selection, and prior MMM results can inform this.\n"],null,["# Refresh the model\n\nRefresh Frequency\n-----------------\n\nModel refreshes can be done as frequently as you would like. Model selection and\ntuning is typically an iterative process, which may need to be refreshed along\nwith new data. You might consider updating the model quarterly, annually, or at\na frequency that matches your marketing budget decision making process.\n\nRecommended: incorporate new data by adding it to the older data\n----------------------------------------------------------------\n\nWe recommend adding the new data to the older data and running Meridian.\nOne ought to consider whether or not to discard the oldest data to accommodate\nthe new data. This may be necessary to stay in the 2-3 year data window that's\ncommon in an MMM. Meridian doesn't model media effectiveness as\ntime-varying. So, the decision to discard old data when appending new data is a\nbias-variance trade-off. A longer time window reduces variance because you have\nmore data, but it can increase bias if media effectiveness and strategies have\nchanged drastically over time.\n\nRecognize that MMM estimates often exhibit high variance. This can mean that\nincorporating even a relatively small amount of new data may have a noticeable\neffect on the model's results. For this reason, there can be valid business\nreasons to set the priors in the new model to encourage the posterior of the new\nmodel to match the posterior of the old model. We recommend that you set priors\nbased on prior knowledge and intuition, and it is reasonable for this intuition\nto be informed by past MMM results. It is your decision as to how strongly you\nwant past MMM results to inform your prior knowledge and intuition. However,\nconsider that setting priors that match an old MMM's results effectively counts\nthe old data twice.\n\nAlternative: model new data disjointly and use priors\n-----------------------------------------------------\n\nSome may consider incorporating new data by fitting a model to just that new\ndata, disjointly from the data used in old models. Although technically\npossible, even for a small amount of data such as a quarter, this is generally\nnot recommended.\n\nModeling the new data completely disjointly from the old data won't properly\nconsider lagged effects. Meridian allows media data to include more (older) time\nperiods than the KPI and controls data. This allows the lagged effects to be\nmore accurately modeled beginning with the first time period of KPI data. It is\nbest to include max_lag time periods of media data prior to the first time\nperiod of KPI data whenever possible.\n\nA small amount of new data is likely not informative enough for the model to\nmake conclusions (see Amount of data needed). One may want to incorporate the\ninformation from the old data by using a prior distribution informed by the\nposterior of the older model. While the full joint posterior distribution of all\nparameters theoretically contains all information from older data (and using it\nas a prior for new data would be similar to fitting a new model that combines\nboth old and new data), Meridian uses independent prior distributions for\nindividual parameters. Therefore, even if the posterior distribution for each\nindividual parameter were carried over as its prior, it might not fully capture\nthe complete joint posterior distribution, which accounts for interdependencies\nbetween parameters. Additionally, Bayesian models require a parametric prior\ndistribution for each parameter. MCMC sampling provides an empirical sample from\nthe posterior, which may or may not have a suitable parametric approximation for\ndirect use as a prior."]]