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!
投资回报率、边际投资回报率和响应曲线
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。
增量效果
对于给定的媒体渠道 \(q\),增量效果定义为:
\[\text{IncrementalOutcome}_q = \text{IncrementalOutcome} \left(\Bigl\{
x_{g,t,i}^{[M]} \Bigr\}, \Bigl\{ x_{g,t,i}^{[M](0,q)} \Bigr\} \right)\]
其中:
- \(\left\{ x_{g,t,i}^{[M]} \right\}\) 是观测到的媒体值
- \(\left\{ x_{g,t,i}^{[M] (0,q)} \right\}\) 表示所有渠道的观测媒体值,但渠道 \(q\)除外(该渠道在所有位置的观测媒体值均设置为零)。更具体地说:
- \(x_{g,t,q}^{[M] (0,q)}=0\ \forall\ g,t\)
- \(x_{g,t,i}^{[M](0,q)}=x_{g,t,i}^{[M]}\ \forall\ g,t,i \neq q\)
投资回报率
渠道 \(q\) 的投资回报率定义为:
\[\text{ROI}_q = \dfrac{\text{IncrementalOutcome}_q}{\text{Cost}_q}\]
其中, \(\text{Cost}_q= \sum\limits _{g,t} \overset \sim x^{[M]}_{g,t,q}\)
请注意,投资回报率的分母表示特定时间段内的媒体费用,该时间段与定义增量效果的时间段一致。因此,分子中的增量效果包括在此时间窗口之前执行的媒体的滞后效应,同样,也不包括在此时间窗口期间执行的媒体的未来效应。这样一来,分子中的增量效果与分母中的费用并不完全一致。不过,从相对较长的时间窗口来看,这种不一致就不那么明显了。
需要注意的是,反事实媒体情景 (\(\left\{ x_{g,t,i}^{[M](0,q)}
\right\}\)) 可能实际上并没有在数据中体现出来。在这种情况下,有必要根据模型假设进行外推,以推理出反事实情景。
响应曲线
根据增量效果的定义,渠道 \(q\) 的响应曲线被定义为一个函数,该函数以渠道 \(q\)支出函数的形式返回增量效果:
\[\text{IncrementalOutcome}_q (\omega \cdot \text{Cost}_q) =
\text{IncrementalOutcome} \left(\left\{ x^{[M](\omega,q)}_{g,t,i} \right\},
\left\{ x^{[M](0,q)}_{g,t,i} \right\}\right)\]
其中, \(\left\{ x^{[M](\omega,q)}_{g,t,i} \right\}\) 表示所有渠道的观测媒体值,但渠道 \(q\)除外(该渠道在所有位置的观测媒体值均需乘以系数 \(\omega\) )。更具体地说:
- \(x^{[M](\omega,q)}_{g,t,q}=\omega \cdot x^{[M]}_{g,t,q}\ \forall\ g,t\)
- \(x^{[M](\omega,q)}_{g,t,i}=x^{[M]}_{g,t,i} \forall\ g,t,i \neq q\)
边际投资回报率 (mROI)
渠道 \(q\) 的边际投资回报率 (mROI) 定义为:
$$
\text{mROI}_q = \left(\dfrac{1}{\delta \cdot \text{Cost}_q} \right) \text{IncrementalOutcome} \left( \left\{ x^{[M](1+\delta,q)}_{g,t,i} \right\},
\left\{x^{[M](1,q)}_{g,t,i}\right\} \right)
$$
其中, \(\delta\) 是一个较小的量,例如 \(0.01\)。
请注意,响应曲线和边际投资回报率的定义隐含了一个假设,即每个媒体单位的费用保持不变,始终等于每个媒体单位的历史平均费用。
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):2025-07-30。
[null,null,["最后更新时间 (UTC):2025-07-30。"],[[["\u003cp\u003eIncremental outcome measures the change in outcome attributed to a specific media channel by comparing observed media values to a scenario where that channel's values are zero.\u003c/p\u003e\n"],["\u003cp\u003eROI is calculated by dividing the incremental outcome of a media channel by its cost, reflecting the return on investment for that channel.\u003c/p\u003e\n"],["\u003cp\u003eResponse curves illustrate the relationship between media spend on a specific channel and the resulting incremental outcome, providing insights into channel effectiveness at different investment levels.\u003c/p\u003e\n"],["\u003cp\u003eMarginal ROI measures the incremental outcome gained by increasing spend on a specific channel by a small percentage, indicating the return on additional investment in that channel.\u003c/p\u003e\n"],["\u003cp\u003eThese metrics rely on counterfactual scenarios, sometimes requiring model-based extrapolation when observed data doesn't fully represent those scenarios.\u003c/p\u003e\n"]]],["Incremental outcome for a media channel is calculated by comparing observed media values to a scenario where that channel's values are zeroed out. ROI is the incremental outcome divided by the channel's cost. Response curves show how incremental outcome changes with varying spend on a channel. Marginal ROI (mROI) measures the change in incremental outcome from a small increase in channel spend, assuming a constant cost per media unit. Counterfactual scenarios where channels are zeroed out might need to be inferred by the models.\n"],null,["This section covers the key metrics of Meridian - Return on Investment (ROI),\nmarginal ROI (mROI) and response curves.\n\nQuick Takeaways\n\nIncremental Outcome, return on Investment (ROI), marginal ROI (mROI), and\nresponse curves are the tools that turn your model's findings into actionable\nbusiness strategy. They help you answer the most critical marketing questions:\n\"How well did my channels perform?\" and \"Where should I spend my next dollar?\"\n\nBy understanding these metrics, you can identify your most efficient channels,\nunderstand current saturation levels, and optimize your budget to maximize\nyour business outcomes. Response curves, in particular, provide a powerful\nvisualization of how the incremental outcome responds to more spending, which is\nthe foundation of data-driven budget allocation.\n\nMarketing Example\n\nImagine you run an online shoe store. You spend \\\\$10,000 on a video media\nchannel. After running your Meridian model, you find that the channel\ncaused \\\\$25,000 in incremental sales.\n\n- **Incremental Outcome** is the value your marketing caused. For example, your total sales were \\\\$150,000, but Meridian estimates that without the campaign, sales would have been \\\\$125,000. The Incremental Outcome is the difference: \\\\$25,000.\n- Your **ROI** is \\\\$2.50 (\\\\$25,000 sales / \\\\$10,000 cost), meaning you earned \\\\$2.50 for every dollar spent. (See [Considerations for interpreting ROI\n and response curves](#considerations_for_interpreting_roi_and_response_curves) for more details on how this is calculated).\n- The **response curve** shows you how your sales would change at different spend levels. It shows that spending even more money yields progressively smaller returns.\n- Your **mROI** is the return you'd get from some small increase in spend (for example, the next dollar). If your channel is nearing saturation, the mROI might be only \\\\$0.80, signaling it's time to invest elsewhere.\n\nRule-of-Thumb Recommendation\n\n- **Use ROI to evaluate historical performance**: It gives you a clear, overall grade for how effective your past spending was on a given channel.\n- **Use response curves to optimize future budgets**: They visualize the point of diminishing returns, helping you understand how much you can invest in a channel before it becomes inefficient.\n- **Use mROI to evaluate saturation level**: If the mROI is much lower compared to the ROI, then the channel is beginning to saturate at historical spend level. Channels with the highest mROI are the best for investing additional funds.\n\nComparison Table\n\n| Metric | Best For | Definition |\n|---------------------|--------------------------------------------------------------|---------------------------------------------|\n| **ROI** | Evaluating past performance. | A historical, channel-wide average. |\n| **Response Curves** | Optimizing future spend and visualizing diminishing returns. | Incremental outcome as a function of spend. |\n| **mROI** | Understanding current saturation level. | The return on the next dollar spent. |\n\nCode Examples\n\nRefer to [Example - ROI, mROI \\& Response Curves in Meridian](/meridian/notebook/ROI_mROI_Response_Curves)\nfor working code examples.\n\nDetailed Explanation\n\nThis section provides a deeper dive into the definitions and methodologies\nbehind ROI, mROI, and response curves.\n\nIncremental Outcome Explained\n\nThe foundation for ROI, mROI, and response curves is **incremental outcome** .\nThis is the portion of your outcome, such as sales or\nconversions, that was caused by a specific marketing activity.\nMeridian calculates this by comparing the actual outcome to a **counterfactual**\nscenario where the marketing activity never happened.\nFor paid media, the incremental outcome can be further contextualized by\nits spend in the following ways:\n\n- The response curve estimates the incremental outcome at any given spend level.\n- ROI is the incremental outcome at your historical spend level divided by the spend.\n- mROI is the incremental outcome on your next dollar spent above the historical budget level.\n\nHow Response Curves Are Generated\n\nA response curve visualizes the relationship between spend and incremental\noutcome for a single channel, assuming all other channels' spending remains\nthe same.\n\nMeridian generates this curve at different spending levels for a\nchannel. It scales the channel's historical spend up or down by a factor\n(for example, from 1.2x the historical spend) and estimates the incremental\noutcome at each level. The historical distribution of spending over time and\ngeography (the **flighting pattern**) is preserved\nduring this scaling. This process reveals the point at which a channel becomes\nsaturated and further investment yields diminishing returns.\n\nConsiderations for interpreting ROI and response curves\n\n- **Lagged effects**: The ROI definition uses a channel's total cost over a specific period as the denominator. The numerator is the incremental outcome accrued during that same period. This numerator includes lagged effects from ads that ran before the period but excludes future effects from ads that ran during the period. Over a long time window (for example, one year), this has a minor effect on the ROI estimate. However, for shorter periods, the effect can be more meaningful.\n- **Extrapolation risk**: Calculating the incremental outcome requires the model to estimate what would have happened if spend was zero. If you have always spent consistently on a channel, the model has little data for this zero-spend scenario and must extrapolate based on its learned assumptions. Extrapolation risk also affects incremental outcome estimation for points on the response curve that are greater than historical spend, and the risk increases the further out you go.\n\nMathematical Appendix\n\nThis section contains the mathematical underpinnings behind ROI, mROI and\nresponse curves.\n\nIncremental outcome\n\nFor a given treatment variable \\\\(q\\\\), the incremental outcome is defined as:\n\n\\\\\\[\\\\text{IncrementalOutcome}_q = \\\\text{IncrementalOutcome} \\\\left(\\\\Bigl\\\\{\nx_{g,t,i}\\^{\\[M\\]} \\\\Bigr\\\\}, \\\\Bigl\\\\{ x_{g,t,i}\\^{\\[M\\](0,q)} \\\\Bigr\\\\} \\\\right)\\\\\\]\n\nWhere:\n\n- The function $\\\\text{IncrementalOutcome}()$ is defined [here](/meridian/docs/basics/incremental-outcome-definition) and is a more generic function that represents the incremental outcome between any two media counterfactual scenarios (not necessarily the incremental outcome of one isolated treatment variable).\n- \\\\(\\\\left\\\\{ x_{g,t,i}\\^{\\[M\\]} \\\\right\\\\}\\\\) are the observed treatment values\n- \\\\(\\\\left\\\\{ x_{g,t,i}\\^{\\[M\\] (0,q)} \\\\right\\\\}\\\\) denotes the observed treatment values for all treatments except treatment \\\\(q\\\\), which is set to its baseline value \\\\(b_q\\\\) everywhere. More specifically:\n - \\\\(x_{g,t,q}\\^{\\[M\\] (0,q)}=b_q\\\\ \\\\forall\\\\ g,t\\\\)\n - \\\\(x_{g,t,i}\\^{\\[M\\](0,q)}=x_{g,t,i}\\^{\\[M\\]}\\\\ \\\\forall\\\\ g,t,i \\\\neq q\\\\)\n\nFor paid and organic media, the baseline values \\\\(b_q\\\\) are zero. For\nnon-media treatment variables, the baseline value can be set to the observed\nminimum value of the variable (default), the maximum, or a user-provided float.\n\nROI\n\nThe ROI of channel \\\\(q\\\\) is defined as:\n\n\\\\\\[\\\\text{ROI}_q = \\\\dfrac{\\\\text{IncrementalOutcome}_q}{\\\\text{Cost}_q}\\\\\\]\n\nWhere \\\\(\\\\text{Cost}_q= \\\\sum\\\\limits _{g,t} \\\\overset \\\\sim x\\^{\\[M\\]}_{g,t,q}\\\\)\n\nNote that the ROI denominator represents media cost over a specified time period\nthat aligns with the time period over which the incremental outcome is defined.\nAs a result, the incremental outcome in the numerator includes the lagged effect\nof media executed prior to this time window, and similarly excludes the future\neffect of media executed during this time window. So, the incremental outcome in\nthe numerator does not perfectly align with the cost in the denominator.\nHowever, this misalignment will be less material over a reasonably long time\nwindow.\n\nNote that the counterfactual media scenario (\\\\(\\\\left\\\\{ x_{g,t,i}\\^{\\[M\\](0,q)}\n\\\\right\\\\}\\\\)) may not actually be represented in the data. When this happens,\nextrapolation based on model assumptions is necessary to infer the\ncounterfactual.\n\nResponse curves\n\nGeneralizing the incremental outcome definition, the response curve is defined\nfor channel \\\\(q\\\\) as a function which returns the incremental outcome as a\nfunction of the spend on channel \\\\(q\\\\):\n\n\\\\\\[\\\\text{IncrementalOutcome}_q (\\\\omega \\\\cdot \\\\text{Cost}_q) =\n\\\\text{IncrementalOutcome} \\\\left(\\\\left\\\\{ x\\^{\\[M\\](\\\\omega,q)}_{g,t,i} \\\\right\\\\},\n\\\\left\\\\{ x\\^{\\[M\\](0,q)}_{g,t,i} \\\\right\\\\}\\\\right)\\\\\\]\n\nWhere \\\\(\\\\left\\\\{ x\\^{\\[M\\](\\\\omega,q)}_{g,t,i} \\\\right\\\\}\\\\) denotes the observed\nmedia values for all channels except channel \\\\(q\\\\), which is multiplied by a\nfactor of \\\\(\\\\omega\\\\) everywhere. More specifically:\n\n- \\\\(x\\^{\\[M\\](\\\\omega,q)}_{g,t,q}=\\\\omega \\\\cdot x\\^{\\[M\\]}_{g,t,q}\\\\ \\\\forall\\\\ g,t\\\\)\n- \\\\(x\\^{\\[M\\](\\\\omega,q)}_{g,t,i}=x\\^{\\[M\\]}_{g,t,i} \\\\forall\\\\ g,t,i \\\\neq q\\\\)\n\nMarginal ROI (mROI)\n\nThe marginal ROI (mROI) of channel \\\\(q\\\\) is defined as: \n$$ \\\\text{mROI}_q = \\\\left(\\\\dfrac{1}{\\\\delta \\\\cdot \\\\text{Cost}_q} \\\\right) \\\\text{IncrementalOutcome} \\\\left( \\\\left\\\\{ x\\^{\\[M\\](1+\\\\delta,q)}_{g,t,i} \\\\right\\\\}, \\\\left\\\\{x\\^{\\[M\\](1,q)}_{g,t,i}\\\\right\\\\} \\\\right) $$\n\nWhere \\\\(\\\\delta\\\\) is a small quantity, such as \\\\(0.01\\\\).\n\nNote that the response curve and marginal ROI definitions implicitly assumes a\nconstant cost per media unit that equals the historical average cost per media\nunit."]]