Comparing in-house vs. outsourcing of MMM (Multi-Marketing Memory): Which tasks should be kept in-house in the open-source era?

Rather than viewing MMM (Marketing Mix Modeling) in-house as a binary choice of "do everything in-house or outsource everything," it's more practical to divide MMM into six tasks and decide whether to handle each task in-house or outsource it to an external partner. With the release of open-source MMMs such as Google's Meridian and Meta's Robyn, it's now possible to create models without licensing fees. On the other hand, the people and time required for the remaining tasks—defining the analysis question, gathering data, interpreting the results, obtaining internal agreement, and adjusting budget allocations—have not changed significantly.

In-house development of MMM means taking on some of the tasks related to MMM within the company.

In-house development of MMM means designing, building, and operating the MMM analysis within your own company's team, rather than outsourcing the analysis. In practice, this may involve handling only a part of the process internally, such as preparing data in-house while collaborating with external partners to build the models. Therefore, instead of viewing it as an either/or choice between in-house development and outsourcing, it's necessary to carefully consider which tasks should be handled internally.

The basic mechanism of MMM is as follows:What is MMM? An explanation of the basics, analysis procedures, and implementation points of Marketing Mix Modeling. .

In open-source MMM, the cost has decreased in model building.

The emergence of open-source MMMs has primarily reduced costs in the area of ​​model building. As of October 2026, the following three are representative examples:

• Meridian (Google):This MMM library, based on Bayesian statistics (a statistical approach that incorporates past knowledge and experimental results as assumptions and then revises them to suit the data), became available to everyone in January 2025. It supports model correction using experimental results (reflecting results from experiments conducted in different regions into the model) and budget allocation simulations. In 2026, a screen that allows users to compare allocation scenarios without writing code (Scenario Planner) and a library for conducting experiments in different regions (GeoX) will be added.
• Robyn (Meta):This MMM library combines a regression analysis method (ridge regression) that can stably estimate the effects of multiple advertisements even when they overlap, with an algorithm that automatically searches for the optimal settings, and includes functions for correction based on experimental results and budget allocation.
• PyMC-Marketing (PyMC Labs):This is an MMM library released by the developers of PyMC, a Bayesian statistics library for Python, and is provided under the Apache 2.0 license, which allows commercial use.

Since all of these libraries are available free of charge, you no longer need to allocate a software budget to create models. However, the libraries do not make decisions on your behalf regarding which variables to include, what assumptions to make, or how to interpret the results. Also, in September 2026, Google announced a mechanism for an AI agent to assist with data quality checks and model building for Meridian users, so some aspects of data preparation may change in the future. Please note that the functionality and development status of each library are subject to change, so please check the latest information on the official website before considering using them.

The reason why open-source MMMs have become widespread is,MMM's Latest Trends and Future Outlook (2024 Edition)But I am introducing it.

There are companies in Japan that are working on developing MMM (Multiple Memory Mixing) systems in-house.

In Japan, companies that have internalized MMM (Market Milling Management) have also made their efforts public. In a technical blog published in May 2025, the eyewear brand JINS cited the high cost of outsourcing, the lack of transparency in the analysis, and the time it took to obtain results as reasons for internalization. On the other hand, they also revealed that they faced challenges in data quality, handling the residual effect (the part of the effect that remains afterward) and saturation (when the effect plateaus even if advertising is increased), designing experiments in different regions, and coordinating between data science and marketing teams. Nevertheless, they concluded that internalization allowed them to increase the frequency and speed of report updates (Source: JINS Tech Blog ""Implementing and Solving Challenges of In-House MMM Production").

Thus, in-house development of MMM is becoming a realistic option for domestic companies. However, the cost reduction is mainly for model building, and the personnel and time required for other tasks remain. Therefore, when considering in-house development, it is practical to start by breaking down the MMM tasks into their individual components.

The six tasks of MMM: Whether to handle them in-house or outsource them is decided on a task-by-task basis.

The process of using MMM (Model-Based Modeling) to revise budget allocation involves six steps: question design, data preparation, model building and updating, interpretation, agreement, and implementation. Whether to handle these steps in-house or outsource them to external partners can be decided for each of these six steps.

The following table summarizes the overview of six tasks and how the emergence of open-source MMM and AI has changed them.

Question design: Which budget decisions will change based on this analysis?

Question design is the process of deciding, before the analysis, which budget decisions will be based on the results of MMM (Market Management Modeling). For example, "revise the allocation between television and digital advertising for the next fiscal year," "adjust the allocation of digital advertising quarterly," and "decide whether or not to try new media" will all require different data periods and different model update frequencies.

Conditions for handling this internally:This task is handled internally, whether it's done in-house or outsourced. Only the company knows which decisions are made, when, and under whose authority.

Situations where external partners are helpful:This is where you need to organize questions into an analyzable format and identify questions that are difficult to answer with MMM. MMM is not good at showing the effectiveness of measures that have never been implemented or measures with very small investment amounts, or the order in which individual customers were exposed to advertisements. Also, since MMM is an analysis for reviewing allocations over several months to a year, if you are adjusting the cost per acquisition on a daily basis, measurement and verification on the advertising side will provide answers more quickly.

Data organization: Data that is separated between the company and agencies is consolidated on a weekly basis.

Data preparation involves gathering the data necessary for MMM (Market Management Model), organizing it into a consistent unit (such as weekly), and maintaining it in a usable state each time the model is updated. This includes performance metrics such as sales and conversions, marketing activity data such as TV GRP and ad impressions, and other factor data such as promotions, weather, seasonality, and competitor trends. A minimum of 13 months of data is needed to validate the model in light of seasonal fluctuations, and 2-3 years of data allows for more reliable analysis.

Before collecting data, clarifying the following four points will help you determine the scope of the data you need.

• Number of products/brands to analyze
• The form of sales (store, e-commerce, or both) and the form of distribution (direct sales, wholesale/retail).
• The magnitude of seasonal fluctuations in sales
• The importance of offline media such as television and outdoor advertising.

In some cases, sales data is kept internally, while data on television and digital advertising placements is kept with advertising agencies. In such cases, since the units and formats differ for each medium, it is necessary to standardize the weekly aggregation of television GRP and digital advertising impressions, and to pre-determine the classification of media with the agency.

Conditions for handling this internally:You'll need data engineers who can continuously process data, and coordinators who can receive data updates from other departments and agencies. Data preparation is required every time the model is updated.

Situations where external partners are helpful:This is the initial dataset creation. If you decide on the classification and aggregation methods for each medium from the beginning, subsequent updates will be easier to manage within the company.

Model building and updating: Open source, most unusual tasks

Model building and updating involves creating a model by deciding which variables to include and what assumptions to make, and then regularly updating it with new data. The biggest change in open-source MMMs is this process, which now allows the use of Bayesian statistics and experimental corrections (such as incorporating results from experiments conducted in different regions into the model) without incurring licensing fees.

Nevertheless, there are still choices to be made within the model. Depending on the assumptions made about how long the effects of advertising last (residual effect) and at what point the effect plateaus as advertising spending increases (saturation), the conclusions for budget allocation can change even with the same data. The experimental results used for correction can similarly affect the conclusions.

How to combine multi-stage effects is also a crucial aspect of determining the model structure. To see indirect effects, such as TV commercials increasing searches, which in turn increase sales, or promotions boosting distribution (the number of stores carrying the product) and leading to increased sales, the model structure needs to be tailored to the business. Meridian expanded its support to include non-advertising variables such as price and promotions in 2025, and to brand-related metrics in 2026, but how to combine multi-stage effects is ultimately up to the user.

Conditions for handling this internally:Rather than simply using the library's default settings, you need analysts who can explain the reasoning behind each assumption. If you update the model quarterly or monthly, you also need a system in place to continue that work.

Situations where external partners are helpful:This is necessary when verifying the model design during startup, or when structures beyond the initial setup are required, such as in cases of multi-stage effects. XICA's MMM uses path analysis (patented) to model the process from awareness to purchase in multiple stages, thereby capturing these indirect effects.

Interpretation: When the results differ from expectations, how far back should we go to revise them?

Interpretation involves determining whether the results of MMM are genuine discoveries or errors in the model, and deciding how far back to revise the model if the results differ from expectations. At XICA, we believe this is the task that requires the most skilled personnel when developing in-house.

The reason is that even if a model fits historical data well, it doesn't necessarily mean that its effect estimations are correct. The Meridian documentation explains that directly verifying whether the causal estimation of how much advertising increased results is correct requires a well-designed experiment, which is difficult, so indirect metrics are usually relied upon. It also states that metrics such as how well the model can predict periods not used for training should not be the primary way to evaluate a model (Source: Google Meridian documentation)Assess the model fit and the results)).

For example, in a brand that concentrates its TV commercials during the month of a new product launch, the effect of the launch itself and the effect of the advertising coincide. Even if the model fits sales trends well, it may be overestimating part of the growth due to the launch as being due to advertising. These errors can be found by comparing the results with those of experiments conducted in different regions, checking sales during periods when advertising was stopped, and comparing them with business events that occurred around the same time, such as expanding distribution.

Conditions for handling this internally:We need someone who understands both the statistical assumptions and the business situation, and who can determine how far back to review the data, assumptions, and model structure when the results differ from expectations. Knowledge of statistics alone or business alone makes it easy to overlook plausible errors.

Situations where external partners are helpful:This is when unexpected results occur, leading to major budget decisions. By reviewing the assumptions and data from a different perspective before incorporating them into the budget, you can reduce the risk of changing allocations based on incorrect conclusions.

Agreement: Can management, finance, and agencies communicate based on the same figures?

Agreement is the process of creating a situation where the same figures can be used to discuss the MMM results with management, the finance department, agencies, and other parties who need to be convinced to accept the changes in allocation. No matter how accurate the analysis may be, the actual budget allocation will only change once those figures are accepted within the company.

What makes reaching an agreement difficult is the relationship between those who measure the effectiveness and those who spend the budget. The neutrality of the results is particularly likely to be questioned in the following three situations:

• When the marketing department's analytics team evaluates their own department's budget:When evaluating the effectiveness of a budget within the same department, management planning and finance departments may ask whether the results are independent of those evaluations.
• When an advertising agency handles both ad placement and performance measurement:When the party proposing and purchasing the advertising space is the same as the party measuring its effectiveness, it becomes easier to perceive that there is room for vested interests to influence the results.
• When using MMM published by the advertising platform:Meridian is published by Google, and Robyn by Meta, and both include ads they sell in their evaluation criteria. While the code is publicly available and can be examined, the US industry media outlet AdExchanger has reported concerns about the large number of advertisers using Meridian with its default settings (July 14, 2026) (Source: AdExchanger)Google's Meridian And Meta's Robyn: A Gift To Measurement Or Trojan Horses?)).

Conditions for handling this internallyThe analysis team should be independent of the departments with the budgets being evaluated, or there should be separate procedures for verifying the results. Furthermore, pre-determining with management planning and finance which metrics to use and what to judge will reduce the time spent re-examining the validity of the figures after the results are released.

Situations where external partners are helpful:This is necessary when the results affect a large budget or when there are differing opinions within the company regarding how the results will be received. Results verified by a third party with no vested interest in the budget being evaluated serve as justification for explanations within the company. On the other hand, if the analysis team is independent of the department that manages the budget and the results are accepted within the company, this need is less pronounced.

Execution: Change the allocation and verify the results in the next analysis.

Execution involves incorporating the results of MMM into policy plans, adjusting budget allocations, and verifying how those changes affected results in the next model update. If the allocation isn't changed, the analysis will simply be an explanation of past performance.

Conditions for handling this internally:Since budget allocation is always decided internally, it's essential that someone with the authority to change the allocation is involved in the MMM discussion. Furthermore, in companies with regular personnel changes, the reasoning behind past allocation decisions is easily lost when the person in charge changes. Keeping a record of the reasoning behind decisions allows successors to review the model's results and connect them to past decisions.

Situations where external partners are helpful:This is when translating the analysis results into allocation proposals and policy plans for each media type, and when continuously iterating through the process from question design to implementation each time the model is updated.

Thus, open-source MMM has primarily changed the model building and updating aspects of the six tasks, while the remaining tasks need to be decided by each company, with each company having to determine who will handle what. Depending on how these tasks are divided, the following three combinations will emerge.

In-house, hybrid, and outsourced: Three combinations seen through six ways of dividing tasks.

The distinction between in-house development, a combination of in-house and outsourcing, and outsourcing depends on which of the six tasks is primarily handled internally. The three most common combinations are primarily outsourcing, a combination of in-house and outsourcing, but in all combinations, the design and execution of the problem are primarily handled internally.

The following table shows a typical example of work distribution; actual work distribution varies from company to company.

When outsourcing is the primary approach is suitable.

Suitable situations:This is useful when there are no in-house analysts and reliable results are needed first, or when the results are to be used in major budget decisions and therefore require verification by a third party.

Internal organizational structure required:This role involves defining the research question, providing data, and determining allocation based on the results. It can be done even if there is no in-house model builder.

Signs that things are not going well:There is no one within the company who can question the results. If we simply accept external reports without question, there will be no opportunity to review the situation when the results differ from expectations.

When combined use is appropriate

Suitable situations:This situation arises when you have in-house analysts who want to be able to explain the results internally, but you still need external expertise for model design and validation.

Internal organizational structure required:We have analysts who can work with models in collaboration with external partners, and a system in place to continuously maintain and improve data.

Signs that things are not going well:The responsibility for building and updating the model is unclear—whether it lies internally or externally. This can lead to delays in updates and failure to record changes in assumptions.

When an in-house approach is suitable

Suitable situations:This applies when you have the necessary personnel to handle data and models, when you want to update the models frequently, or when there are reasons why you cannot send data outside the company. In Japan, there are companies like JINS that have chosen to internalize the process due to the cost of outsourcing and the time it takes to get results, and have made their efforts public.

Internal organizational structure required:This involves data engineers, analysts who can explain the reasoning behind each assumption, people who can interpret results in a business context, and alternative procedures for verifying those results.

Signs that things are not going well:Currently, there's only one person who understands the model. If that person leaves due to a transfer or retirement, the model will stop being updated, and no one will be able to explain the assumptions behind the results.

The combination can be changed along the way. For example, you could spend the first year building the model in collaboration with an external partner, and then, once your internal analysts can explain the assumptions, you could move the model building and updating to be done in-house. Which combination works best depends on what tasks your company is currently capable of handling.

Decision-making process: Which tasks should our company handle internally?

You can determine which tasks your company should handle internally by assessing whether you are currently able to handle each of the six tasks in-house. Once you've decided which tasks to outsource, narrow down your options by considering the type of partner and asking questions to clarify before signing a contract.

For each of the six tasks, we will determine whether it can be handled in-house.

Tasks for which you can answer "yes" to the "verify" question in the following table are likely to be handled in-house. Tasks for which you can answer "no" will require the help of an external partner or the establishment of an internal system.

The reason for verifying the question design first is that the scope of necessary data and the frequency of model updates can only be determined once the question is decided.

What differs depending on the external party: the form of the contract and the compensation structure.

External partners can be compared based on two axes: "the form of the contract and the fee structure" and "what the compensation is for." The tasks requested in the contract and the way the partner earns compensation do not always coincide, and this discrepancy affects the neutrality of the results and how model updates are handled.

The following table shows typical trends, but actual contracts vary depending on the company and the project.

Five questions to consider when choosing an external partner

When choosing an external partner, you can assess the transparency of their model, how they verify results, their neutrality, and what remains after the contract is signed by how they answer the following five questions.

The fourth question is about avoiding dependence on external partners. Even when outsourcing, confirming in advance what will remain in-house after the contract is signed will leave you with options when considering bringing the work in-house or changing partners in the future.

How XICA is involved as an external partner

XICA is a data science firm founded in 2012, and has been offering MMM as a consulting service called "MAGELLAN" since 2016, utilizing its proprietary tools. Consultants and data scientists are involved in six tasks, from question design and data preparation to model building and updating, interpretation, agreement, and execution. XICA may also be involved with companies that are already using MMM internally, or companies that are using models from other companies, by verifying the results as a third party during the interpretation and agreement phases.

For more details, please see the MMM service page.For those already using MMM."

FAQ

Would bringing MMM (Model-Making Memory) in-house reduce costs?

While software costs may decrease, the total cost may not necessarily. Open-source MMM eliminates the licensing fees for building and updating models, but the tasks of data preparation, interpretation, and agreement still require in-house personnel and time. The more tasks are moved in-house, the more personnel are needed, so the total cost will vary depending on which tasks are handled internally. Incidentally, JINS, which developed its own MMM in-house, cited cost reduction as one of the benefits of in-house development in a technical blog published in 2025.

What kind of personnel are needed to internalize MMM (Multi-Marketing Memory) development?

The necessary personnel depend on which of the six tasks will be handled in-house. If data preparation is done in-house, a data engineer and someone to coordinate with agencies will be needed. If model building and updating are also done in-house, an analyst who can explain the reasoning behind each assumption will be added. If interpretation is also done in-house, someone with knowledge of both statistics and business will be required. In any case, someone with budget allocation authority needs to be involved in the design and execution of the research question.

If we want to switch from outsourcing our MMM (Marketing Management) to doing it in-house, where should we start?

First, we start by confirming whether we can continue to receive advertising data from the agency. This is because even after internalizing the process, much of the advertising data for television and digital advertising will still be obtained through the agency. Then, initially, we focus on one brand and one performance indicator, and for a certain period, we compare the results of the existing model and the internal model side by side to understand the reasons for the differences and facilitate the transition.

How much data and time is needed to internalize MMM (Marketing Memory) development?

At least 13 months of data are required, and 2-3 years of data will yield more reliable results. The timeframe varies greatly depending on the state of the data and the company's internal structure. For reference, when XICA performs MMM, it usually takes about 3 months until the first report is released (this varies depending on the data preparation status). In the case of in-house development, this may be further complicated by the need to arrange data transfer and allow time for analysts to become familiar with the libraries.

Summary

When developing an MMM in-house, the key decision is which of the six tasks will be handled internally. Open-source MMMs reduce the cost of creating models, but the people and time needed for question design, data preparation, interpretation, consensus, and execution remain. Paying an external partner is worthwhile for tasks that cannot yet be handled internally or tasks that should not be handled solely internally.

For more detailed information on how to implement MMM, please see the following article.

• What is MMM? An explanation of the basics, analysis procedures, and implementation points of Marketing Mix Modeling.
• “Modern MMM (Marketing Mix Modeling)” Practical Guide for Marketers
• 10 Common Misconceptions and Truths about MMM: Is it Really Only for High-Priced, Large Corporations?
• Bringing marketing analysis in-house | Explaining the advantages and disadvantages

Author information

XICA Co., Ltd. Marketing Team