Customer Lifetime Value Modeling

Customer Lifetime Value Modeling (Probabilistic / BTYD)

Overview

Customer Lifetime Value (CLV) is the present value of the future cash flows attributed to a customer relationship. This skill covers the probabilistic “buy-till-you-die” (BTYD) family — statistical models that decompose CLV into (1) how often a customer transacts while active, (2) whether/when they silently churn, and (3) how much they spend per transaction — then discount the expected future stream to present value.

Two orthogonal axes define the model landscape (Fader/Hardie taxonomy):

Choosing the wrong quadrant is the #1 modeling error. Subscriptions/SaaS are contractual (you see the cancellation) → sBG / survival. Retail, e-commerce, donations are non-contractual (you infer churn) → Pareto/NBD family.

Authoritative source corpus: Bruce Hardie’s notes (brucehardie.com), the Fader/Hardie/Lee Marketing Science papers, and the three reference implementations — lifetimes (Python, archived), CLVTools (R), and PyMC-Marketing (Python, Bayesian, the active successor).

Core Concepts

1. The buy-till-you-die (BTYD) framework

A customer is “alive” until an unobserved dropout, transacting stochastically while alive. Models pair a counting process (transactions while alive) with a timing process (lifetime/dropout), each with cross-customer heterogeneity. First introduced by Schmittlein, Morrison & Colombo, “Counting Your Customers: Who Are They and What Will They Do Next?”, Management Science 33(1):1–24 (1987) (https://pubsonline.informs.org/doi/10.1287/mnsc.33.1.1). Lineage: Retina.ai “History of BTYD” (2023).

2. Pareto/NBD

The original non-contractual continuous-time model. NBD (Poisson–gamma mixture) for transaction counts while alive; Pareto (exponential–gamma mixture) for the unobserved lifetime. Four parameters (r, α, s, β). Powerful but numerically awkward (Gaussian hypergeometric functions), which motivated BG/NBD. (Schmittlein et al. 1987; CLVTools pnbd; PyMC-Marketing Pareto/NBD notebook.)

3. BG/NBD (“Counting Your Customers the Easy Way”)

The workhorse. Replaces Pareto’s continuous dropout with a beta-geometric story: a customer flips a coin to churn immediately after each transaction (prob. p, beta-distributed across customers); active counts are NBD. Far easier to fit (estimable in Excel), nearly identical predictive accuracy. Fader, Hardie & Lee, Marketing Science 24(2):275–284 (2005) (http://brucehardie.com/papers/018/fader_et_al_mksc_05.pdf). Quirk: in BG/NBD a customer cannot churn until after their first repeat purchase, so it understates one-and-done customers — which MBG/NBD fixes.

4. MBG/NBD (Modified BG/NBD)

Adds a dropout opportunity at time zero (right after the first purchase), so customers who never repeat can be “dead”. Expected-repeat estimates nearly match BG/NBD, but alive/dead classification of zero-repeat customers is more realistic. Batislam, Denizel & Filiztekin, IJRM 24(3) (2007); implemented as ModifiedBetaGeoModel.

5. Gamma-Gamma monetary model

Separately models spend per transaction (frequency models only predict counts). Assumptions: (a) value varies randomly around the customer’s mean; (b) mean spend varies across customers but not over time; (c) spend is independent of the transaction process — verify frequency and monetary value are roughly uncorrelated before trusting it. Fit only on repeat purchasers. Fader, Hardie & Lee, “RFM and CLV: Using Iso-Value Curves”, JMR 42(4):415–430 (2005) (https://www.brucehardie.com/papers/rfm_clv_2005-02-16.pdf).

6. RFM as model inputs (sufficient statistics)

BTYD models need only per-customer Recency, Frequency, and “T” — R and F are sufficient statistics for the likelihood. Conventions (easy to get wrong):

• frequency = number of repeat purchases (total − 1).
• recency = time between first and last purchase (NOT time since last purchase, the marketing-RFM convention).
• T = customer “age” = first purchase to end of observation.
• monetary_value = average value of repeat transactions.

7. Discounted Expected Residual Transactions (DERT) → CLV

CLV (non-contractual) = (expected spend from Gamma-Gamma) × DERT, where DERT is the present value of all expected future transactions discounted to the end of the calibration period (integral from T to ∞). Use a continuously-compounded discount rate (e.g. 15%/yr ≈ 0.0027/week). Fader/Hardie originally called this DET. (RFM-CLV 2005; CLVTools pnbd_DERT; Fader/Hardie note 033.)

8. sBG — shifted-beta-geometric (contractual / discrete churn)

Subscriptions/contractual settings: each period a customer renews with prob. θ or cancels with 1−θ; θ is fixed per customer, beta-distributed across the base. Projects observed retention into a full survival curve and explains the observed rise in aggregate retention over time as a heterogeneity sorting effect, not behavior change. Fader & Hardie, “How to Project Customer Retention”, J. Interactive Marketing 21(1):76–90 (2007); extended in “Customer-Base Valuation in a Contractual Setting”, Marketing Science 29(1):85–93 (2010).

9. BG/BB — discrete-time non-contractual

Discrete-time analog of Pareto/NBD: transactions per period are Bernoulli (buy/no-buy) instead of Poisson, paired with a beta-geometric dropout — for “transaction opportunities” data (annual donations, periodic catalog buyers). Closed-form. Fader, Hardie & Shang, Marketing Science 29(6):1086–1108 (2010); lifetimes BetaGeoBetaBinomFitter.

10. Predictive vs. historical CLV

Historical CLV sums realized past margin (backward-looking). Predictive CLV forecasts future value via models (BTYD, ML, or naive ARPU/churn). The naive ARPU ÷ churn shortcut assumes a single constant retention rate — biased low when retention is heterogeneous (Fader/Hardie 2010). Prefer model-based predictive CLV with uncertainty intervals.

11. Cohort-based CLV

Group customers by acquisition period and track value per cohort. Reveals retention dynamics and acquisition-quality drift a base-wide average masks; pairs with sBG on multicohort data. (Keep retention-curve fitting itself in da-34; here it is a CLV input/segmentation lens.)

12. CAC:LTV ratio (unit economics)

LTV:CAC measures payback on acquisition spend. Rules of thumb: ~3:1 healthy target (B2C SaaS ≈ 2.5:1, B2B SaaS ≈ 4:1); below 2:1 = unsustainable; above ~5:1 = likely under-investing. CAC payback: healthy 6–12 months, elite < 3 months. Use a margin-based, discounted predictive LTV — gross-revenue LTV inflates the ratio.

Tools / Frameworks

Methodology (non-contractual continuous: common case)

1. Confirm the quadrant. Non-contractual + continuous → proceed. Contractual → sBG/survival. Discrete opportunities → BG/BB.
2. Build RFM summary (rfm_summary() / lifetimes summary_data_from_transaction_data). Watch the recency definition.
3. Fit a frequency/dropout model (BG/NBD default; MBG/NBD if many one-and-done; Pareto/NBD as benchmark).
4. Check the model: holdout calibration, tracking plot, P(alive) distribution.
5. Fit Gamma-Gamma on repeat purchasers; first verify low corr(frequency, monetary).
6. Compute discounted CLV = E[spend] × DERT over a finite horizon, continuously-compounded discount rate.
7. Validate on a holdout window by RFM decile.
8. Segment / act: rank by predicted CLV and P(alive); feed CAC:LTV.

Anti-Patterns

• Wrong quadrant (Pareto/NBD on a subscription business, or sBG on e-commerce).
• Marketing-RFM recency (“days since last purchase” instead of “first-to-last span”) — silent severe bias.
• Gamma-Gamma without the independence check.
• Naive ARPU ÷ churn as ground truth — biased low under heterogeneity.
• Un-discounted / infinite-horizon CLV — inflates value and LTV:CAC.
• Fitting Gamma-Gamma on all customers instead of repeat purchasers only.
• Trusting lifetimes for new long-lived projects — it’s archived.

Troubleshooting

• P(alive) implausibly high for everyone → BG/NBD with many one-and-done customers; switch to MBG/NBD.
• Optimizer fails / NaN log-likelihood (Pareto/NBD) → numerical instability in hypergeometric terms; use log-sum-exp-patched BTYD or BG/NBD.
• Gamma-Gamma returns absurd spend → filter to frequency > 0; use average repeat value, not total.
• Holdout over-predicted → calibration window caught a promo spike; re-split or model seasonality outside BTYD.
• CLV explodes → infinite horizon or zero discount rate; cap horizon, set continuously-compounded rate.

References

1. Schmittlein, Morrison & Colombo, Management Science 33(1):1–24 (1987) — https://pubsonline.informs.org/doi/10.1287/mnsc.33.1.1
2. Fader, Hardie & Lee, BG/NBD, Marketing Science 24(2):275–284 (2005) — http://brucehardie.com/papers/018/fader_et_al_mksc_05.pdf
3. Fader, Hardie & Lee, RFM and CLV / Gamma-Gamma + DERT, JMR 42(4):415–430 (2005) — https://www.brucehardie.com/papers/rfm_clv_2005-02-16.pdf
4. Fader & Hardie, Gamma-Gamma note 025 — https://www.brucehardie.com/notes/025/gamma_gamma.pdf
5. Fader & Hardie, sBG / “How to Project Customer Retention”, J. Interactive Marketing 21(1):76–90 (2007)
6. Fader & Hardie, “Customer-Base Valuation in a Contractual Setting”, Marketing Science 29(1):85–93 (2010) — http://brucehardie.com/papers/022/fader_hardie_mksc_10.pdf
7. Fader, Hardie & Shang, BG/BB, Marketing Science 29(6):1086–1108 (2010) — http://www.brucehardie.com/papers/020/fader_et_al_mksc_10.pdf
8. Batislam, Denizel & Filiztekin, MBG/NBD, IJRM 24(3) (2007)
9. Fader & Hardie, “What’s Wrong With This CLV Formula?” note 033 — http://www.brucehardie.com/notes/033/what_is_wrong_with_this_CLV_formula.pdf
10. PyMC-Marketing CLV docs (v0.15.x, 2024–2025) — https://www.pymc-marketing.io/en/stable/notebooks/clv/clv_quickstart.html
11. CLVTools (R) — https://www.clvtools.com/
12. lifetimes (Python, archived) — https://github.com/CamDavidsonPilon/lifetimes
13. Phoenix Strategy Group, LTV:CAC SaaS benchmarks — https://www.phoenixstrategy.group/blog/ltvcac-ratio-saas-benchmarks-and-insights