Decision Packs | TriaCaf

Marketing Measurement

Marketing Mix Modeling

Which channels are actually driving revenue, and what happens to that revenue if you move budget between them.

PyMC-Marketing MMM with adstock and saturation curves per channel, weakly-informative priors, and posterior predictive checks. Fans out across multiple adstock specifications and reports ROAS as a posterior distribution, not a point estimate — budget elasticity comes out the same run.

Multi-Touch Attribution

Which channels deserve credit across the full customer journey — not just whichever one got the last click.

Markov-chain attribution over customer touchpoint paths, compared against first-touch and last-touch baselines. Per-channel shares sum to 100%; the L1 distance between the data-driven and heuristic allocations tells you which channels are actually load-bearing in the model choice, not just in the story.

Synthetic Control

Build a credible "what would have happened without this" for a single market, region, or launch.

Abadie-style synthetic control: constructs a weighted counterfactual from a donor pool of untreated time series. Three specs — unconstrained weights, Dirichlet-constrained classical SCM weights, and a permutation placebo test — so a treatment claim with a placebo p-value above 0.2 gets flagged, not asserted.

Geo Test Designer

Design a real geo experiment before you spend the budget — know your detection limit going in.

Matches control geos to a treated geo on pre-period correlation, recommends an assignment, and runs a simulation power analysis to report the minimum detectable lift at 80% power. Deterministic design statistics, not a model fit — this pack tells you whether a test is worth running before it starts.

Customer Economics

Customer Lifetime Value

What each customer, or customer segment, is really worth over their lifetime — not just their first order.

Bayesian hierarchical Pareto/NBD-lite model: Poisson–Gamma pooling on transaction frequency, a recency-based alive-probability, and Gamma-distributed order value. Reports portfolio LTV and per-customer forecasts with 95% credible intervals across three shrinkage specs, so you can see whether the pooling assumption is doing the work.

Customer Churn (Survival)

Which customers are actually at risk of leaving, and roughly when — before it shows up in the revenue numbers.

Bayesian Weibull survival analysis across three specs — pooled, per-segment hazard curves, and Weibull proportional hazards with covariates. Reports median time-to-churn with credible intervals, hazard ratios, and flags any customer whose posterior survival probability at six months drops below 0.5.

Price Elasticity

How much a price change would actually move demand and revenue — and where the elastic/inelastic line sits.

Bayesian log-log demand model — log(Q) = α + β·log(P) + γ·controls + ε — across three specs (static, plus controls, plus regional hierarchy) so seasonality or region heterogeneity shifting the elasticity estimate is visible, not hidden. Reports the elasticity posterior plus a counterfactual: quantity and revenue response to a 10% price move.

Forecasting & Monitoring

Hierarchical Demand Forecasting

How much of each SKU to expect next period — including the slow-moving, intermittent items most forecasters get wrong.

Bayesian hierarchical Poisson model with partial pooling across category × region. Three specs — no-pooling MLE baseline, hierarchical Poisson with per-category shrinkage, and zero-inflated Poisson for intermittent demand — report per-SKU λ posteriors and flag SKUs where more than 30% of periods are zero and the specs disagree.

Bayesian Time-Series Forecast

Where a metric is headed next, with an honest range attached — not a single confident-looking line.

Trend, Fourier seasonality, and AR(1) residual terms, fit across three specs of increasing complexity with the dominant period auto-detected via FFT. Reports per-spec in-sample RMSE and multi-step-ahead posterior predictive intervals — divergence across specs tells you whether seasonality is actually load-bearing.

Bayesian Anomaly Detection

Catch the data points that don't fit — and know which flagged anomalies are trustworthy versus noise.

Residual-based anomaly scoring across three specs — naive Z-score, Bayesian rolling moving-mean with a Gamma-Poisson conjugate, and seasonal-adjusted residual scoring. Reports per-observation scores and the cross-spec overlap (Jaccard index): anomalies flagged by all three specs are trustworthy, spec-divergent ones are candidates only.

Causal & Experimentation

Causal Inference

Did a specific change — a launch, a policy shift, a campaign — actually cause the result, or would it have happened anyway?

A Bayesian interrupted time series: trend, seasonality and outside drivers (spend, price, promotions) are modelled around the date of the change, and the lift is reported as a full posterior distribution rather than a single number — so the uncertainty in "what would have happened" is visible, not smoothed over. Three specifications show whether the answer depends on those modelling choices; for a single treated region or store with untreated peers, the Synthetic Control pack is the stronger design.

Bayesian A/B Testing

Whether a test result is real, and how confident you should actually be — not just whether p is under 0.05.

Beta-binomial multi-arm analyzer across three specs — non-hierarchical (Jeffreys prior), hierarchical with partial pooling across arms, and a Savage-Dickey Bayes-factor test. Reports per-arm lift with 95% credible intervals, P(treatment > control), and Bayes factors — a credible lift is only declared when P > 0.95 or BF₁₀ > 3.

Twelve decision packs.
One verification standard.