Model Conjectures for Team Review

Recommendation status

Everything on this page is a proposal for discussion with the data-modeling team. None of these models is currently trained or implemented in this repository.

The purpose of this page is to preserve focused ideas while colleagues transition out of the project. Each section states what a method does, where it works best, what data it requires, and what the team must decide.

For unfamiliar terms, use the Definitions page.

Decision map

Stage	Candidate	Question it answers	Current data sufficient?
Transparent baselines	First, last, linear, position-based	How do simple rules divide credit?	Yes
Timing baseline	Time decay	Does recent contact deserve more credit?	No; timestamps are missing
Observational path analysis	Markov removal	Which touchpoints matter to path transitions?	Mostly
Interaction-aware allocation	Shapley value	How should joint value be divided fairly?	Yes for synthetic evaluation
Conversion prediction	Regularized logistic model	Which path features predict conversion?	Yes, at aggregated-path level
Event-time modeling	Additive hazard model	How does conversion risk change over time?	No; event times are missing
Causal research	Deconfounded sequence models	Can richer user journeys reduce observational bias?	No
Production causal calibration	Experiment-calibrated ensemble	Which attribution estimates best predict causal lift?	No; experiments are missing
Strategy optimization	Marketing Mix Modeling and constrained allocation	How should future budget change?	Not yet; more history is needed

Level 1: transparent attribution baselines

First-touch attribution

• Method summary: Give 100 percent of conversion credit to the first observed touchpoint.
• Easy example: A shopper sees display → brand search → product search. Display receives all credit.
• Most effective application: Measuring which campaigns introduce people to a brand or product.
• Data needed: Ordered paths.
• Main limitation: It ignores everything that happens after discovery.
• Team decision: Do we want this only as a test baseline, or also as a stakeholder report?

Last-touch attribution

• Method summary: Give 100 percent of credit to the final touchpoint.
• Easy example: In display → brand search → product search, product search receives all credit.
• Most effective application: Simple transaction reporting when the final interaction is the main business concern.
• Data needed: Ordered paths.
• Main limitation: It often overvalues lower-funnel activity that captures demand created elsewhere.
• Team decision: Which existing Amazon reports already use a comparable last-touch definition?

Linear attribution

• Method summary: Divide credit equally among all touchpoints.
• Easy example: Three touchpoints each receive one third of a $90 attributed sale.
• Most effective application: A neutral, easily audited comparison when no evidence supports unequal weights.
• Data needed: Ordered or unordered touchpoint sets.
• Main limitation: Equal credit is an assumption, not a learned result.
• Team decision: Should repeated occurrences of the same touchpoint receive separate shares?

Position-based attribution

• Method summary: Reserve chosen weights for the first and last touchpoints and divide the remainder among the middle.
• Easy example: A three-touch path receives 40 percent, 20 percent, and 40 percent.
• Most effective application: Businesses that explicitly value both discovery and closing activity.
• Data needed: Ordered paths.
• Main limitation: The weights are chosen by people rather than estimated from outcomes.
• Team decision: Which weight settings should become official regression-test fixtures?

Time-decay attribution

• Method summary: Give more credit to touches closer in time to conversion.
• Easy example: An advertisement seen one hour before purchase receives more credit than one seen two weeks earlier.
• Most effective application: Short decision cycles where recency is known to matter.
• Data needed: Real event timestamps and conversion time.
• Main limitation: The current schema has order but no time intervals. Invented timestamps would make the result misleading.
• Team decision: Can Amazon Marketing Cloud provide privacy-safe time buckets that preserve useful elapsed-time information?

Level 2: observational path models

Markov removal-effect attribution

• Method summary: Estimate transition probabilities between path states, then remove one channel and measure how much estimated conversion probability falls.

• Easy example: If removing Sponsored Brands lowers modeled conversion from 10 percent to 8 percent, its removal effect is 2 percentage points.

• Most effective application: Large collections of converting and non-converting paths where sequence matters.

• Data needed: Ordered paths, converted counts, and non-converted counts.

• Current transformation:

  $$\begin{array}{rl}\text{total conversions}& =\text{converted users}\\ \text{total non-conversions}& =\text{users}-\text{converted users}\end{array}$$

• Main limitation: A Markov chain measures observed path structure, not causal incrementality. Results can also depend on the chosen chain order and removal procedure.

• Team decision: Which chain orders and path-pruning rules should be compared?

Shapley-value attribution

• Method summary: Average each touchpoint's marginal contribution across every order in which touchpoints could join a coalition.
• Easy example: Display is weak alone but strong when paired with product search. Shapley allocation shares the partnership value rather than giving it all to the final touch.
• Most effective application: Small or medium numbers of touchpoint categories where interactions matter and a value can be estimated for each coalition.
• Data needed: A trustworthy value function for touchpoint groups.
• Main limitation: Coalition counts grow rapidly, and results are only as credible as the value function.
• Team decision: Should the first implementation use exact enumeration, sampled permutations, or both?

Regularized logistic conversion model

• Method summary: Predict conversion probability from path features while regularization discourages unstable or excessively large coefficients.
• Easy example: Features indicate whether display, Sponsored Brands, and Sponsored Products appeared and in what order. The model estimates conversion probability for each combination.
• Most effective application: Interpretable conversion prediction with many correlated path features.
• Data needed: Both converted and non-converted journeys. Aggregated rows can be trained with user counts as weights.
• Main limitation: Removing a feature in a predictive model creates a counterfactual prediction, not automatically a causal effect.
• Team decision: Which features are allowed: presence, count, order, marketplace, and path length?

Open-source implementations to review

• ChannelAttribution: https://github.com/DavideAltomare/ChannelAttribution
• Multi-Touch Attribution Python package: https://github.com/eeghor/mta

These are reference candidates, not approved dependencies.

Level 3: event-level predictive and causal research

Additive-hazard or time-to-event model

• Method summary: Model how the instantaneous chance of conversion changes over time after advertising events.
• Easy example: A product advertisement may sharply raise purchase likelihood for two days, while a video advertisement may have a smaller but longer effect.
• Most effective application: Timestamped journeys where delay and carryover matter.
• Data needed: Impression or click timestamps, conversion timestamps, and observation-end times for non-converters.
• Main limitation: The current aggregated schema cannot identify event timing.
• Team decision: Is event-time information obtainable at a privacy-safe aggregation level?

CausalMTA-style deconfounding

• Method summary: Learn conversion and attribution while attempting to separate advertising effects from customer preferences that influence which advertisements they receive.
• Easy example: Frequent skincare buyers are both more likely to receive skincare advertisements and more likely to purchase. The model attempts to avoid crediting the entire difference to advertising.
• Most effective application: Rich event-level data with user context and known treatment-selection mechanisms.
• Data needed: Timestamped journeys, non-conversions, stable anonymized identities, and relevant user or context features.
• Main limitation: Hidden confounders can remain, and the current project data is far too aggregated.
• Team decision: Is this a research benchmark or a realistic production candidate?

Causal attention or sequence model

• Method summary: Use a neural sequence model to represent the order and context of many events, then derive touchpoint importance or counterfactual effects.
• Easy example: The meaning of a Sponsored Display touch may differ when it follows video, repeated product search, or a long gap.
• Most effective application: Very large, richly timestamped journey datasets where simpler models have already been benchmarked.
• Data needed: Event-level sequences, context, negative outcomes, and enough volume for held-out testing.
• Main limitation: Complexity does not restore missing causal information and can make errors harder to diagnose.
• Team decision: What measurable improvement over logistic and hazard models would justify this complexity?

Research sources

• CausalMTA: https://arxiv.org/abs/2201.00689
• Causal Attention Model for Multi-Touch Attribution: https://arxiv.org/abs/2012.11403
• Criteo Attribution Modeling for Bidding Dataset: https://ailab.criteo.com/criteo-attribution-modeling-bidding-dataset/

Level 4: production causal calibration

Experiment-calibrated attribution ensemble

• Method summary: Use randomized controlled trials as campaign-level causal targets. Combine predictions from several attribution models, learn which mixture best predicts experimental lift, and then distribute calibrated credit to touchpoints.
• Easy example: Last-touch predicts 1,000 conversions and another model predicts 700, while an experiment estimates 800 incremental conversions. A calibration model learns how to combine and adjust the observational estimates.
• Most effective application: Large advertising systems with a representative, continuing experiment program.
• Data needed: Campaign experiments, touchpoint-level model outputs, conversion histories, and campaign features.
• Main limitation: This repository has no randomized experiment corpus or proprietary shopping signals.
• Team decision: Can the business obtain Amazon multi-touch metrics or run enough holdout tests to create calibration targets?

Industry reference

Amazon's published Multi-Touch Attribution architecture is the principal reference: https://arxiv.org/abs/2508.08209

Level 5: strategy and budget optimization

Marketing Mix Modeling

• Method summary: Marketing Mix Modeling fits aggregate sales over time using channel spend or exposure, seasonality, prices, promotions, carryover, and saturation.
• Easy example: Weekly sales are modeled against Sponsored Products spend, display spend, holidays, and discounts to estimate channel response curves.
• Most effective application: Cross-channel budget planning when user-level tracking is incomplete but long, consistent time series exist.
• Data needed: Preferably two or more years of weekly data, channel inputs, outcomes, controls, and experiment calibration.
• Main limitation: It works at an aggregate level and does not replace path-level attribution.
• Team decision: Do we have enough geographic and temporal variation to identify separate channel effects?

Google Meridian

• Method summary: An open-source Bayesian Marketing Mix Modeling framework with geographic modeling, experiment calibration, reach and frequency support, and budget optimization.
• Easy example: Compare advertising response across regions and use experiment results as prior information.
• Most effective application: Geographic time-series data and teams prepared for Bayesian modeling and heavier computation.
• Main limitation: It is not a path-attribution engine.
• Team decision: Can the team support its data and computing requirements?

PyMC-Marketing

• Method summary: A flexible Bayesian marketing toolkit with configurable carryover, saturation, time-varying effects, calibration, and optimization.
• Easy example: Build a custom model where video has longer carryover than search and quantify uncertainty around each response curve.
• Most effective application: Teams needing model customization and explicit uncertainty.
• Main limitation: Flexibility increases modeling and review responsibility.
• Team decision: Who will own prior selection, diagnostics, and model refreshes?

Meta Robyn

• Method summary: A semi-automated Marketing Mix Modeling framework using ridge regression, time-series decomposition, hyperparameter search, and budget allocation.
• Easy example: Search many plausible carryover and saturation settings, then compare candidate channel decompositions.
• Most effective application: Teams seeking an automated workflow and able to validate multiple candidate models.
• Main limitation: Its repository describes the native Python implementation as beta; extra verification is required.
• Team decision: Is using its R implementation acceptable if Python parity remains uncertain?

Framework sources

• Google Meridian: https://github.com/google/meridian
• PyMC-Marketing: https://github.com/pymc-labs/pymc-marketing
• Meta Robyn: https://github.com/facebookexperimental/Robyn

Recommended training and evaluation contract

1. Use amc_path_report.csv and amazon_ads_daily_touchpoint_performance.csv as model inputs.
2. Never expose simulation_ground_truth.csv to feature construction.
3. Split by time rather than random row so the same scenario does not leak across training and evaluation.
4. Evaluate conversion prediction with weighted log loss and calibration.
5. Evaluate attribution with credit-share error, rank correlation, and purchase and revenue conservation.
6. Evaluate strategy optimization by replaying proposed allocations inside held-out simulator configurations.
7. Repeat across many seeds and deliberately shifted configurations. One seed proves reproducibility, not robustness.
8. Before production, calibrate against randomized experiments or another credible incrementality source.

Questions to take into the modeling meeting

1. Is the immediate objective explanatory attribution, conversion prediction, causal incrementality, or future budget allocation?
2. Can the data source provide non-converting paths, timestamps, and interaction type?
3. What experiment or holdout evidence can serve as real calibration truth?
4. Which simple baseline must every candidate beat?
5. Which errors matter most: wrong total credit, wrong channel ranking, poor probability calibration, or bad budget decisions?
6. Who owns model review after the current team transition?

Continue reading

• Documentation Home
• Definitions
• Dataset Generation Flow
• Peer-Compatible Baseline
• Regional Simulation
• Repository Analysis
• Prompt Record