Modeling firms’ own solicitation behavior alongside customer sequences uncovers rich CLV distributions and identifies customers at a ‘fork’ whose responsiveness makes marketing highly profitable; using the right tail of predicted CLV rather than the mean raises targeting lift (up to 8.3%) while matched comparisons show bimodal customers respond far better to solicitations.

Main Finding

A dual-LSTM—two interacting LSTM networks that jointly predict firm solicitations and customer responses and are simulated forward via Monte Carlo—recovers full customer lifetime value (CLV) distributions (not just means), captures self-reinforcing dynamics (diverging paths and bimodal outcomes), and yields materially better predictive and prescriptive performance than standard CLV models. The model outperforms six benchmarks on 19 of 24 evaluation metrics; ranking by the right tail of the CLV distribution improves targeting lift up to 8.3%; bimodal customers identified by the model show much higher marginal ROI to solicitations (+77%) compared with observably identical unimodal customers (-50%).

Key Points

• Motivation: Most CLV models produce a single expected value per customer, implicitly assuming decision-makers are indifferent to how that value materializes. A survey of 38 senior consultants found 79% would treat two customers with identical expected CLV but different distributional profiles differently.
• Conceptual innovation: Treat CLV as a distribution shaped by the co-evolution of customer behavior and firm strategy (endogenous firm actions), rather than a one-sided stochastic process driven only by customers.
• Model architecture:
  • Two LSTMs: (1) Firm policy LSTM predicts per-period solicitation probability from historical solicitation/donation sequences; (2) Customer behavior LSTM observes the firm-model’s predicted solicitation for the same period and predicts donation probability.
  • Predictions are treated as Bernoulli draws and simulated forward via Monte Carlo hundreds of times to generate CLV distributions (expectation, tails, multimodality).
  • Two-model sequential design prevents negative transfer and enables same-period marketing acceleration effects.
• Empirical results:
  • Data: 9 years of donor solicitation and donation histories for 21,939 individuals.
  • Holdout: 36-month evaluation horizon.
  • Benchmarks: Pareto/NBD, hierarchical Bayesian extensions (Abe), SUR bivariate probit, ElasticNet, single-LSTM, single-transformer.
  • Performance: Dual-LSTM outperforms competitors on 19/24 metrics across seven families of evaluation criteria.
  • Managerial gains: Ranking by the right tail (upside) rather than expected CLV yields up to +8.3% targeting lift.
  • Actionability: Using Coarsened Exact Matching to pair bimodal vs unimodal customers with identical observables, bimodal customers respond substantially more to solicitations (ROI of extra solicitation +77% vs -50% for unimodal); no benchmark model replicates this actionable pattern.
• Intuition shown in data: A homogeneous donor cohort diverged into a bimodal distribution due to firm reactions (fewer solicitations after nonresponse → less future giving), illustrating self-fulfilling, path-dependent dynamics.

Data & Methods

• Dataset: Individual-level fundraising panel (solicitations and donations) over 9 years; N = 21,939 donors.
• Model:
  • Two interacting LSTM networks (firm-policy LSTM first, customer-response LSTM second per period).
  • Input: raw temporal sequences (solicitations, donations, monthly dummies), not aggregated recency/frequency summaries.
  • Sequential Monte Carlo forward simulation: simulate stochastic solicitations and donations period-by-period hundreds of times to obtain a CLV distribution per customer.
• Evaluation:
  • 36-month holdout period.
  • Compared against six benchmarks: Pareto/NBD, hierarchical Bayesian extension, SUR bivariate probit, ElasticNet, single-LSTM, single-transformer.
  • Evaluated on 24 metrics spanning predictive accuracy, distributional fit, and targeting lift (7 metric families).
  • Causal-leaning test for actionability: Coarsened Exact Matching to control observables and compare responsiveness/profitability of bimodal vs unimodal customers to real solicitations.
• Key quantitative results:
  • Outperforms benchmarks on 19/24 metrics.
  • Targeting lift improvement up to +8.3% when ranking by right-tail CLV vs expected CLV.
  • Additional solicitation ROI: +77% for bimodal customers vs −50% for unimodal customers (matched on observables).

Implications for AI Economics

• Modeling endogenous agents: This work shows AI models can and should treat firm policy as an endogenous, stochastic agent when forecasting economic outcomes that arise from agent interaction. Joint simulation of interacting agents (firm and customers) produces richer outcome distributions relevant for decision-making under risk and path dependence.
• Distributional forecasts matter: Point estimates hide actionable risk/upsides. Distributional CLV enables prioritizing customers by upside tail or downside risk (for valuation, resource allocation, incentive design), improving targeting efficiency and revealing segments where interventions have large marginal returns.
• Policy and causal considerations: Learning the firm policy from historical data makes forward projections more realistic than exogenous policy assumptions, but observational learning still faces endogeneity/causal identification limits. Matching and randomized experiments remain important complements to validate causal effects of targeting rules uncovered by the model.
• Transferability & applications: The dual-LSTM framework is applicable to other economic contexts where agent interactions drive outcomes over time (B2B sales, subscriptions, credit/churn, labor markets, platform dynamics). Rich panel data with recorded principal actions is required; computational cost and interpretability trade-offs vs. simpler structural models should be weighed.
• Methodological advance for AI economics: Demonstrates a practical approach—interacting deep-sequence models + Monte Carlo simulation—to recover complex, path-dependent distributions in economic settings. This expands the toolkit for evaluating risk, option-like upside, and strategic interventions in microeconomic decision-making.

If you want, I can: - Extract key figures/tables to visualize the bimodality and simulation outcomes; - Produce a concise checklist for implementing a dual-LSTM CLV pipeline (data requirements, architecture, evaluation); or - Draft targeted experimental designs (A/B or bandit) to causally test the model’s prescriptive recommendations.