Method library
Every method traces to a paper.
No proprietary mystery method. No "our secret algorithm". Open-source primitives, established literature, the Kleyr platform layer wraps them in a productised workflow. Read what's under the hood, decide if you trust it.
Marketing mix modelling
MMM
PyMC-Marketing
open-source · pymc-labs/pymc-marketing
Bayesian MMM backbone. Adstock + Hill saturation, Gaussian-process baselines, hierarchical pooling. Active OSS.
Google Meridian
open-source · 2024 · google/meridian
Hierarchical Bayesian MMM with geo-level pooling (TFP / XLA). Reference for multi-geo customers.
Bayesian MMM, foundational
Jin, Chan, Wang, Sun et al. (Google) · 2017
The reference Bayesian MMM paper. Companion: Chan & Perry on failure modes.
Adstock transforms
various — geometric, Weibull, delayed
Geometric for fast-decay channels, Weibull for flexible shapes, delayed-adstock for slow-ramp media.
Saturation transforms
Hill / Michaelis-Menten / S-curve
Hill is the practical default. Michaelis-Menten as alternate. Logistic for capped contexts.
Stepwise selection
AIC / BIC / p-value criteria
Forward, backward, bidirectional. Surfaced as configurable. Known biases, surfaced honestly.
Causal inference
Experiments
Synthetic Control
Abadie, Diamond & Hainmueller · 2003 / 2010 / 2015
Foundational. The credible counterfactual when classical pre/post DiD breaks down.
Augmented Synthetic Control
Ben-Michael, Feller & Rothstein · 2021
Robustness improvement on classical SCM. What GeoLift uses internally.
Synthetic Difference-in-Differences
Arkhangelsky et al. · 2021
Often outperforms SCM. Optional method.
GeoLift
Meta · open-source · facebookincubator/GeoLift
Python-native end-to-end geo experimentation. v1 default for geo lift.
CausalImpact
Brodersen et al. (Google) · 2015
Bayesian structural time series. Available as a cross-check method via Python port.
CUPED
Deng, Xu, Kohavi & Walker (Microsoft) · 2013
Variance reduction via pre-experiment covariate. Massively reduces required A/B sample sizes.
Causal Forests / GRF
Wager & Athey · 2018
Heterogeneous treatment effects. Advanced mode only — easy to misinterpret.
Double / Debiased ML
Chernozhukov et al. · 2018
ATE estimation with arbitrary ML nuisance models. EconML implementation.
Customer analytics
Customer Knowledge
Pareto/NBD CLV
Schmittlein, Morrison & Colombo · 1987
Foundational BTYD model. Pareto-distributed customer dropout, NBD purchase rate.
BG/NBD
Fader, Hardie & Lee · 2005
Easier to fit than Pareto/NBD, near-identical accuracy. Practical default.
Gamma-Gamma monetary
Fader & Hardie · 2013
Pairs with BG/NBD for revenue forecasts. Independent monetary value model.
HDBSCAN
Campello, Moulavi & Sander · 2013
Density-based hierarchical clustering. No need to choose k, robust to non-spherical clusters.
Survival analysis
Cox · 1972 (proportional hazards)
Time-to-churn modelling. DeepSurv (Katzman et al. 2018) for non-linear hazards.
Calibration
Niculescu-Mizil & Caruana · 2005
Reliability diagrams as the trust signal. Platt scaling and isotonic regression for re-calibration.
Hierarchical reconciliation (MinT)
Hyndman et al. · 2011
Forecasts that agree across aggregation levels. Used by Fast Forwards to keep customer-level and channel-level forecasts coherent.
✦ Get started
Bring your spend, leave with a model you can defend.
Book a 30-minute working session with the Kleyr team. We'll model your last 12 weeks of marketing spend on the call — no slides, no abstract demo.