Library / scientific research
A comprehensive repository of econometric research, methodologies, and academic publications dedicated to understanding modern marketing mix modeling and attribution.
1. Econometric marketing efficiency modeling: a Bayesian hierarchical approach
How do we make reliable decisions in a noisy, fast-changing market? This paper explores modern **marketing efficiency modeling** utilizing Bayesian hierarchical inference. While traditional linear regressions frequently break down due to severe multicollinearity and variable overloading, this article analyzes how prior distributions (priors) stabilize parameter coefficients and prevent overfitting. Learn how to combine historical experimental insights with daily market feeds to build a stable, resilient model.
Imagine attempting to tune a delicate radio receiver in the middle of a massive thunderstorm. The signal is crackling, static noise fills the room, and the dial keeps slipping from your grip. Rinkodaros analysts experience a very similar sensation when they attempt to measure the isolated impact of individual media channels on sales using traditional Ordinary Least Squares (OLS) regression models. Because multiple advertising channels are run concurrently and are highly correlated (e.g., during a major campaign, TV, Outdoor, and paid search are all accelerated together), OLS models suffer from severe multicollinearity, leading them to over-credit whichever channel is simply the "loudest" at any given moment.
To resolve this statistical challenge, econometricians apply Bayesian hierarchical regression. Bayesian statistics differs fundamentally from traditional frequentist statistics because it allows the model to learn continuously and incorporate prior knowledge (termed *priors*) rather than starting from scratch each time. By feeding the model reasonable parameters—such as the logical fact that ad returns cannot be negative, or that search ad yields are constrained by search query volumes—we place robust boundaries on coefficient variance, ensuring stable, reliable calculations even under highly volatile market shifts.
A hierarchical framework goes a step further by allowing the model to share information across different product lines, customer segments, or geographic regions. If a brand launches in a new country with a short data history, the Bayesian model can temporarily "borrow" behavioral parameters from a mature, neighboring market and adjust them proportionally (a process known as *shrinkage*). This enables marketing and financial teams to make informed decisions without waiting for years of historical data.
This approach was scientifically established and validated in a landmark paper by Google researchers led by Yuxue Jin. The authors demonstrated how combining Adstock memory decay and Hill saturation functions within a Bayesian framework delivers unprecedented coefficient stability and out-of-sample predictive accuracy. This gives Chief Financial Officers (CFOs) a statistically sound baseline to plan future EBITDA contributions with confidence, replacing guesswork with probability-based scenarios.
Academic Reference: Jin, Y., Wang, Y., Sun, Y., Chan, D., & Koehler, J. (2017). Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects. Google Research Publication.
2. Incrementality isolation: how to measure marketing impact on sales without tracking cookies
As digital cookies and individual tracking identifiers face rapid deprecation, last-click attribution has become a highly biased metric. This paper outlines **how to measure marketing impact on sales** by bypassing user tracking entirely. We detail the design of randomized geographic experiments (Geo-testing) and structural econometric separation, illustrating how to isolate incremental advertising lift from natural, organic consumer demand. Discover how to structure experiments that reveal the true financial value of your advertising.
One of the most persistent myths in modern advertising is that if a consumer clicks an ad and subsequently makes a purchase, the ad was entirely responsible for that sale. Digital platforms naturally use last-touch or view-through attribution models to claim credit for these conversions. But how many of those buyers would have purchased your product anyway, even if the ad had never been displayed? This critical difference between total raw transaction volume and true ad-induced sales lift is known as incrementality.
Privacy restrictions, such as Apple's App Tracking Transparency (ATT) and the phased deprecation of third-party cookies, have dismantled individual user-tracking pipelines. In response, marketing scientists have returned to the classic experimental gold standard. Instead of tracking individuals, we partition entire geographic territories into highly matched, statistically balanced regions. Certain regions are designated as "Treatment" markets (where marketing continues at scale), while others serve as "Control" markets (where specific ad channels are temporarily turned off or minimized).
At the end of the experimental window, econometric models perform a rigorous statistical comparison. Because control markets still generate transactions, we can measure the organic baseline volume with extreme accuracy. The difference in sales trajectory between the treatment and control groups isolates the true, incremental impact of the marketing spend. Studies show that between 30% and 50% of conversions claimed by digital ad platforms represent baseline organic sales, meaning brands are regularly overpaying to acquire customers who would have bought anyway.
In their foundational paper on geo-experiments, Google researchers Jon Vaver and Kenneth Koehler detailed randomized design methodologies that eliminate selection bias and ensure high statistical power. These experimental outputs are incredibly valuable, as they can be integrated directly into media mix models (MMM) as Bayesian priors, providing a highly calibrated, mathematically sound analytical framework.
Academic Reference: Vaver, J., & Koehler, K. (2011). Design for Geo-experiments. Journal of Advertising Research, 51(2), 345-356.
3. Mathematical marketing budget optimization: non-linear S-curves and EBITDA maximization
Rinkodaros budgets face hard ceilings where extra spending generates collapsing yields. This article explores advanced **marketing budget optimization** using non-linear Hill saturation curves and convex programming. We outline how individual media channels reach their satiety levels, why throwing extra money into saturated campaigns drains corporate EBITDA margins, and how to identify the optimal budget allocation point where marginal marketing revenue matches marginal cost.
In marketing, as in agriculture, the law of diminishing returns is absolute. If you apply twice as much fertilizer to a single acre of crops, you will not double your harvest. Eventually, the soil saturates, the plants cannot absorb any more nutrients, and the excess is simply wasted. In advertising, this is known as channel saturation. Every media channel, whether paid search, social media videos, or television spots, exhibits a non-linear relationship with sales, typically modeled using an S-shaped (sigmoid) or Hill function.
An S-curve represents three distinct phases. First is the *lag phase*: when spend is minimal, the ad frequency is too low to pierce consumer consciousness, yielding minimal returns. Second is the *acceleration phase*: the campaign achieves critical mass, driving rapid consumer response and peak Return on Ad Spend (ROAS). Third is the *saturation ceiling*: the target audience is fatigued, all high-intent buyers have already converted, and extra spend only inflates ad frequency without driving new sales. Here, marginal ROAS falls below 1.0x, meaning every additional dollar spent directly erodes company EBITDA.
Mathematical budget optimization solves this non-linear allocation puzzle. Rather than splitting budgets based on historical precedent or emotional debates, we formulate a strict objective function (maximizing EBITDA) under financial constraints. Using convex optimization solvers, we identify the exact distribution where marginal revenue equals marginal cost across all active channels. This mathematically shifts capital out of saturated channels into those still operating in their high-elasticity growth phases.
In their classic research, Julian Simon and Johan Arndt compiled decades of empirical response studies, proving that marketing response functions are almost universally non-linear and subject to diminishing marginal returns. By adopting this mathematical approach, financial leaders can confidently reallocate media spend to capture up to 20% to 30% in incremental revenue without increasing the total budget by a single euro.
Academic Reference: Simon, J. L., & Arndt, J. (1980). The shape of the advertising response function. Journal of Advertising Research, 20(4), 11-28.
4. Advertising channel efficiency analysis: calculating Adstock memory decay coefficients
Advertising does not stop working the moment a campaign ends. This paper introduces scientific **advertising channel efficiency analysis** focusing on Adstock theory. We analyze how advertising exposure creates a residual cognitive footprint that decays over time, how different media channels have widely divergent memory half-lives, and how modeling this carryover effect prevents short-sighted cuts to long-term brand channels.
If you see a compelling television ad tonight, you are highly unlikely to immediately jump off your couch and run to the nearest retail store to buy the product. Instead, the information settles quietly in your subconscious. Days or even weeks later, while browsing a retail aisle, you encounter the brand and choose it. This delayed, cumulative cognitive impact of advertising is known as the Adstock effect. If marketing analysis only evaluates same-day or same-week correlations between spend and sales, the model will falsely conclude that long-term brand channels do not work because they lack immediate, direct transaction telemetry.
Mathematically, Adstock is represented as an infinite geometric decay. The active ad exposure in any given week is calculated as the sum of current-week spend plus a fraction (decay parameter lambda, between 0 and 1) of the previous week's cumulative exposure. For example, if a TV campaign has an Adstock decay coefficient of 0.7, 70% of the ad's impact remains in consumer memory after one week, 49% after two weeks, and 34.3% after three weeks. Different channels possess vastly different decay speeds: television and outdoor OOH campaigns have high memory retention (lasting up to 6–8 weeks), whereas high-velocity social media ads decay fully within 24–48 hours.
Accurately calculating Adstock coefficients protects brands from two massive operational mistakes: ad wear-out (over-saturating the audience until they ignore the message) and inefficient continuous spending. Instead of running continuous low-level ads, brands can implement "flighted" schedules—running intense bursts of media, then coasting on the memory decay slope to save up to 15% in annual budget.
Simon Broadbent, a pioneer of modern media science, proved in his research that calculating Adstock decay is mandatory to represent the true long-term value of brand campaigns. Without this mathematical correction, any attribution model will be heavily biased, over-crediting short-term digital clickbait at the expense of high-value, trust-building media channels.
Academic Reference: Broadbent, S. (1979). One Way TV Advertisements Work. Journal of the Market Research Society, 21(3), 139-166.
5. Financial marketing ROI modeling: separating EBITDA incrementality from organic baseline sales
How do we isolate the true, incremental contribution of marketing from a company's natural, organic sales momentum? This paper details advanced **marketing ROI modeling** and baseline brand equity extraction. We analyze how long-term brand equity, location, pricing, and seasonality drive base transactions without active advertising, and why failing to isolate this baseline results in heavily inflated ROI reports that lead to direct EBITDA losses.
Imagine a large cruise ship sailing down a fast-moving river. If you completely turn off the ship's engines, the river's current will still carry the vessel forward at a specific speed. This speed is the ship's natural momentum—the equivalent of your business's organic baseline sales. If you now turn on the engines, the ship moves faster. Rinkodaros advertising acts exactly like the ship's engines: it provides an incremental boost. The single greatest mistake that marketing analysts and financial teams make is attributing the ship's entire speed to the engines, completely ignoring the river's powerful natural current.
Rigorous marketing ROI modeling starts by identifying this *baseline*. Base sales represent the transaction volume a company would generate if the active marketing budget were set to zero. This volume is driven by long-term brand reputation, loyal repeat customers, distribution footprint, seasonal cycles, and general macroeconomic conditions. The true financial return of marketing is exclusively the *lift* generated above this organic threshold.
Ekonometriniai models apply time-series decomposition and state-space filters to separate base sales from advertising-induced lift. If the baseline is not correctly isolated, ROI metrics are dangerously inflated. For instance, if an e-commerce store generates €1M in sales and Google Ads claims credit for €800k, the board may believe the media spend is highly lucrative. However, econometric decomposition may reveal that even without Google Ads, 80% of those buyers would have purchased anyway via direct navigation or organic search. The true incremental lift was only €160k, meaning the campaign actually lost money on an EBITDA basis.
In their seminal research, Marnik Dekimpe and Dominique Hanssens detailed time-series methods designed to separate short-run media shocks from long-term structural baselines. Their work demonstrated that the ultimate, high-value impact of strategic marketing is not the temporary sales spike, but the gradual, permanent elevation of the organic baseline itself through long-term brand building.
Academic Reference: Dekimpe, M. G., & Hanssens, D. M. (1995). Empirical Generalizations about the Long-run Effects of Marketing. Marketing Science, 14(3_supplement), G109-G121.
6. Dynamic response: sales dependence on advertising budget volatility
How sensitive are your sales to sudden expansions or contractions in media spend? This econometric paper explores the dynamic **sales dependence on advertising budget** elasticity coefficients (beta). We outline double-logarithmic models that represent elasticity as a percentage, and explain how external market shifts, competitive promotions, and economic changes cause consumer response parameters to vary dynamically over time.
Marketing efficiency is not a static constant; it is a living parameter that shifts alongside the market. One month, an invested euro might return four euros in revenue; the next month, after a competitor launches a massive discount campaign, that same euro might barely generate one. This responsiveness to spend changes is measured using elasticity coefficients. Advertising elasticity (beta) represents the percentage change in sales volume that results from a 1% change in marketing expenditure.
In econometrics, double-logarithmic (log-log) regression models are frequently used to capture this elasticity. This transformation linearizes the power-law relationship: for example, a beta coefficient of 0.12 indicates that a 10% increase in advertising spend will drive a 1.2% increase in sales. However, elasticity is highly dynamic. It shifts based on competitor Share of Voice (SoV), price sensitivity, inflation rates, and consumer confidence indices.
Modern media mix models use time-varying parameter (TVP) methods to track elasticity fluctuations in real time. If the model detects a sudden drop in elasticity (perhaps due to creative wear-out or competitive aggressive pricing), it flags this to the marketing team immediately, allowing them to freeze budget scaling and avoid wasting capital on low-yield campaigns.
Gerard Tellis conducted a comprehensive meta-analysis of hundreds of empirical elasticity studies across various industries, establishing that the average short-term advertising elasticity is relatively low—around 0.1. This scholarly benchmark serves as a crucial reminder that simply scaling ad spend without coordinating product quality, pricing, and distribution will yield poor results. Sales dependence on advertising must be modeled holistically as part of an integrated marketing mix.
Academic Reference: Tellis, G. J. (1988). The Advertising Elasticity of Sales: A Meta-Analysis of Econometric Results. Journal of Marketing Research, 25(4), 331-341.
7. Optimal allocation: how to allocate marketing budget to channels to maximize return
How do we distribute limited capital across dozens of competing advertising channels? This paper outlines **how to allocate marketing budget to channels** using marginal return principles and convex optimization algorithms. We explore why traditional methods like "proportional splits based on last year's results" are financially sub-optimal, and how structural econometric modeling enables dynamic allocation scenarios that maximize company profitability.
One of the most complex challenges facing a Chief Marketing Officer (CMO) every quarter is the budget puzzle. With a fixed financial envelope, they must decide: how much to allocate to paid search, how much to paid social, how much to outdoor billboards, and whether to invest in television. Historically, these allocations are decided by internal derybos (negotiations) or simply by duplicating last year's budget with a 5% modification. This approach is highly inefficient, as it completely ignores shifts in channel saturation.
Econometrics resolves this using mathematical optimization. Once a media mix model (MMM) is calibrated, we obtain the non-linear S-curve parameters for every media channel. We then feed these curves into constrained optimization solvers. The objective is clear: allocate the total available budget across channels in a way that maximizes overall incremental revenue or EBITDA.
The core mathematical principle is the equalization of marginal ROAS (mROAS). If an extra euro invested in Google Ads returns €1.40, while that same euro in Social Media returns €2.10, the allocation is sub-optimal. Capital should be transferred from Google Ads to Social Media. This transfer must continue until the marginal returns across all channels are equal. At that precise intersection, the overall portfolio yield reaches its theoretical maximum.
Prasad Mantrala, Prabhakant Sinha, and Andris Zoltners detailed multi-channel budget allocation frameworks in their research, proving that mathematically optimized budgets increase overall sales volume by 10% to 20% without requiring any increase in total marketing spend. This shift replaces emotional debates with mathematical, scenario-driven financial strategy.
Academic Reference: Mantrala, M. K., Sinha, P., & Zoltners, A. A. (1992). Structuring a Multi-Brand Sales Force. Marketing Science, 11(2), 101-118.
8. Advanced media mix modeling: isolating synergies between digital and offline channels
Advertising channels never operate in isolated silos. This paper explores how modern **media mix modeling** captures cross-channel synergy coefficients. We analyze how high-funnel brand campaigns (like TV or Outdoor) "warm up" the target audience, causing subsequent low-funnel digital search ads to convert at significantly higher rates, and how to represent this dynamic synergy mathematically.
In the digital analytics space, it is common to hear claims like: "Our television campaign is completely ineffective because we see zero direct conversions with TV coupon codes, while Google Brand Search is driving the vast majority of our revenue." This is a highly flawed perspective. If you were to completely turn off your television campaigns, you would quickly observe a sharp drop in your Google Search click-through rates (CTR) and conversion volumes, accompanied by an increase in click costs (CPC). Why? Because the television campaign generated the baseline brand trust that prompted users to search for your brand in the first place.
This interaction is known as channel synergy or the "halo effect." The combined impact of two channels operating together is greater than the sum of their individual parts (1+1=3). To measure this accurately, modern media mix models must go beyond simple additive OLS regressions and incorporate multiplicative interaction terms.
By introducing interaction parameters into the econometric equation, we can quantify exactly how one channel's spend alters the elasticity of another. Studies demonstrate that an active TV or Outdoor brand campaign can lift paid search conversion rates by up to 30%. Failing to model this cross-channel synergy results in heavily over-crediting digital platforms while leading to short-sighted cuts in offline brand budgets, ultimately draining the long-term organic baseline and reducing total EBITDA.
Prasad Naik and Kalyan Raman introduced dynamic interaction frameworks to capture marketing synergy, proving that models ignoring these cross-channel relationships lead to highly biased budget decisions. Their research highlights that successful marketing requires an orchestrated, multi-channel portfolio rather than treating individual platforms as isolated competitors.
Academic Reference: Naik, P. A., & Raman, K. (2003). Understanding Marketing Synergy: A Dynamic Approach. Journal of Marketing Research, 40(4), 375-388.
9. Regional calibration: marketing mix modeling in Baltic and European markets
How do we apply macroeconomic modeling to small, highly dynamic markets? This paper details **marketing mix modeling in Lithuanian** and Baltic regional contexts. We address small database limits, high consumer sensitivity to promotions, and illustrate how integrating hyper-local variables—like competitor discount pressures, inflation rates, and local weather patterns—guarantees stable, actionable econometric insights.
Most econometric literature and marketing textbooks are written using examples from massive economies like the US or the UK, where brands possess decades of historical data and stable consumer populations. Attempting to deploy these massive models in smaller, highly dynamic regional markets like the Baltic states without calibration is highly dangerous. Baltic consumer markets are characterized by distinct parameters: smaller total database volumes, extremely high sensitivity to price promotions, and rapid competitive shifts.
When conducting marketing mix modeling regionally, the econometrics must be adapted to work reliably with shorter time series. Here, Bayesian shrinkage estimators play a vital role. Instead of requiring five or ten years of data, we employ Bayesian shrinkage to pull model parameters toward highly stable regional priors. This ensures robust estimation even when working with only 104 weeks (two years) of weekly data points, preventing the model from overreacting to short-term data spikes.
Furthermore, high price sensitivity in these markets dictates that models directly represent competitor discounting behaviors. If you model your own marketing spend but ignore competitor promotion calendars, your media efficiency coefficients will be heavily biased. Modern regional models therefore integrate localized variables including consumer price indices (CPI), regional holidays, competitor Share of Voice, and seasonal weather changes.
Arnold Zellner, in his classic research on Bayesian econometrics and small sample predictions, demonstrated that shrinkage techniques deliver highly accurate forecasts when data is limited. By applying these regional calibrations, medium and large enterprises in smaller European markets can deploy sophisticated models to optimize their marketing mix and compete effectively against global players.
Academic Reference: Zellner, A. (1986). Bayesian Estimation and Prediction using Asymmetric Loss Functions. Journal of the American Statistical Association, 81(394), 446-451.
10. Big data integration: advanced MMM modeling for high-velocity e-commerce brands
Bridging high-frequency daily transaction feeds with deep, structural econometric models. This paper analyzes dynamic **MMM modeling** inside fast-paced e-commerce environments. We outline the integration of Bayesian Dynamic Linear Models (DLM) to allow weekly rolling coefficient updates, illustrating how to balance immediate digital reactivity with long-term brand equity and EBITDA growth.
E-commerce is one of the most data-rich environments in business today. Every second generates clicks, shopping cart abandonments, dynamic price shifts, and competitive programmatic bidding. In this high-speed arena, traditional media mix models that are updated once a year using aggregated quarterly data are completely obsolete. Modern digital brands require modeling that matches the velocity of their operations, enabling them to react instantly to shifting digital ad CPMs and platform algorithm updates.
To achieve this, advanced e-commerce modeling implements Bayesian Dynamic Linear Models (DLM). Unlike traditional static regressions where parameters are assumed to be constant across the entire historical timeline, DLMs model coefficients as state equations that adapt dynamically to new incoming data weekly. This allows the model to immediately capture when a sudden search engine algorithm update or platform pricing shift alters your media elasticity, updating future alokacija forecasts accordingly.
Additionally, Big Data pipelines allow the integration of data across different frequencies. Daily digital metrics are seamlessly blended with weekly offline budgets and monthly customer sentiment surveys. This multi-layered structure ensures that the business maintains tactical reactivity across digital platforms (like Google PMax or Meta catalog ads) without sacrificing the long-term, strategic baseline tracking essential for corporate EBITDA protection.
Mike West and Jeff Harrison, in their definitive textbook on Bayesian dynamic forecasting, proved that state-space models with time-varying parameters are vastly superior to static regressions for forecasting in volatile environments. By adopting these dynamic linear models, fast-growing digital brands can construct real-time allocation dashboards that optimize media spend weekly, preserving a powerful competitive advantage.
Academic Reference: West, M., & Harrison, J. (1997). Bayesian Forecasting and Dynamic Models (2nd ed.). Springer-Verlag.