Contracts That Pay Only for What the Data Can Identify: Age–Period–Cohort Indexing of Buy-Back Obligations

When a supplier promises to take back or compensate for units that fail, it ends up carrying three kinds of risk mixed together. The first is wear that grows with a unit's age. The second is quality tied to the batch or year in which the unit was made, which the supplier controls. The third is calendar-time shocks, such as a recession or an energy crisis, which hit every unit in the market at once and which the supplier does not control. Coordination contracts in operations research price this obligation as a single random liability. If the supplier dislikes risk rather than caring only about averages, it charges a premium for all of that variance, including the shocks it cannot influence. This proposal writes a buy-back contract whose payments are adjusted, after the fact, by an estimate of the market-wide calendar-time shock. The supplier then bears the risks it controls and is insured against the one it does not. Cost-sharing and revenue-sharing studies show that removing uncontrollable variance shrinks risk premia (Qian et al.; Luo et al.), so that part of the idea is not new. The new part is a statistical obstacle that standard contract theory never meets.

The obstacle is that the three effects cannot all be recovered from data. Cohort equals period minus age, so a linear trend can be moved freely among the three components without changing the fit. That ambiguous trend is exactly what the parties would argue about: is failure rising because the supplier's batches are getting worse, or because the macro environment is drifting? Our central observation is that the parts of the period effect that are identified are its bends and breaks, meaning deviations from a straight line. Those are precisely what joinpoint regression estimates. Joinpoint regression fits a piecewise-linear trend and locates the points where its slope changes. We propose contracts that index payments only to this identified part and leave the unidentifiable linear drift as an explicit, negotiated escalator clause. We will test whether this recovers most of the value of full indexing while staying immune to the arbitrary assumptions that full indexing requires.

Core Claim

Write the log-odds of a unit failing as $h(a,p,c) = f_A(a) + f_P(p) + f_C(c)$. Here $a$ is the unit's age, $p$ is the calendar period, and $c = p - a$ is the production cohort. Because $a - p + c = 0$, the transformation $f_A \to f_A + \gamma a$, $f_P \to f_P - \gamma p$, $f_C \to f_C + \gamma c$ leaves $h$ unchanged for every slope $\gamma$. The period curve's slope is therefore arbitrary. Its deviations from any straight line, written $\tilde f_P(p)$, are not arbitrary: they are identified without any extra assumption. The proposed contract charges the supplier, per period, its realized failures minus its exposure-weighted expected excess failures from $\tilde f_P$, plus a contractual escalator $\rho\,p$. The escalator $\rho$ is bargained over in the open rather than buried inside an identifying constraint:

$$ L_i(p) = F_i(p) - E_i(p)\,\big[\sigma(\hat h_{-i}(p)) - \sigma(\hat h_{-i}(p) - \hat{\tilde f}_{P,-i}(p))\big] + \rho\,p . $$

In this expression, $L_i(p)$ is supplier $i$'s liability in period $p$, $F_i(p)$ is its realized failures, $E_i(p)$ is its units at risk, and $\sigma$ is the logistic function. The subscript $-i$ means the period estimate is fitted on all other suppliers' units. A single supplier therefore cannot move its own index. This resolves the worry in the research file that retrospective trend analysis and forward-looking mechanism design are at odds. The contract fixes the formula ex ante and applies it ex post, at claim settlement, when the joinpoint and age–period–cohort estimates are available.

The claim has three testable parts. First, under risk neutrality, indexing changes only who pays whom and adds no efficiency, so the reference result is a gain of zero. Second, under mean-variance or Conditional Value-at-Risk preferences (CVaR, the average loss in the worst tail), the gain in certainty-equivalent surplus grows with the agent's risk aversion and with the share of liability variance coming from period curvature. This follows because the mean-variance risk premium is proportional to the variance borne. Third, indexing on the identified component is robust: transfers under full-period indexing swing materially as the identifying constraint changes, and transfers under our contract do not.

Research Program

The work has three pieces.

The first is analytical. We extend the single-period Stackelberg buy-back model to multiple periods. In this model the leader commits to contract terms and the follower then responds. Following Yang et al., who show that buy-back and revenue-sharing contracts coordinate a retailer that evaluates outcomes by CVaR, we add a failure hazard with age, period and cohort structure. The supplier's effort shifts only $f_C$. We then derive the risk premium under three regimes: no indexing, full period indexing, and identified-only indexing. Two quantities matter. One is the premium left over from the linear period drift that identified-only indexing does not cover. The other is any incentive distortion, which arises because a misestimated index can leak some cohort risk away from the supplier. Cai et al. show that a joint revenue-and-cost-sharing contract cannot coordinate perfectly when both parties are risk-averse, and that it works only when $R_s < 2R_r$, where $R$ is risk tolerance. We will check whether removing the common-shock variance widens that region. Our prediction is that it should, because the conflict between the two parties' risk preferences operates mainly on variance neither of them controls.

The second piece is estimation. We fit the period component with the neural discrete-time survival model of Wang et al., which models default hazards by loan age, calendar time and origination vintage. That paper breaks the identification deadlock by fitting the period effect to unemployment and house prices. We drop that anchor when building the index and keep only the detrended period curve. We then fit a joinpoint model to that curve, choosing the number of knots by the Bayesian Information Criterion (BIC). The joinpoint fit compresses the index into a few auditable slope changes, in the style of the combined joinpoint and age–period–cohort analysis of Chen et al.. The research file notes that standard joinpoint tests assume independent errors, which operational data violate. We therefore use block-permutation tests.

The third piece is evaluation of the contract on real hazards.

The main testbed is the Freddie Mac loan-level data used by Wang et al.: over one million loans originated from 2004 to 2013, with monthly performance records. It fits this question unusually well, because a loan originator's obligation to repurchase defective loans is literally a buy-back contract. Originators control underwriting quality, which is the cohort effect, and the 2009 crisis is a textbook period shock. The secondary testbed is a synthetic battery warranty chain. Age and cohort degradation curves come from BatteryML, and we inject period shocks. The setting is the new-energy-vehicle recycling chain of Chen et al., whose authors flag their single-period assumption as the main limitation.

Baselines

The do-nothing reference is the unindexed buy-back contract, the coordinating contract of Yang et al., extended across periods so that the supplier bears all realized failures. We run it on the same simulated and replayed hazard paths. We also run it under risk neutrality, where every contract should tie, as a sanity check.

The strongest existing competitor is full period indexing using the macro-anchored period effect of Wang et al.. That paper is the only work in the file that both separates period from cohort and ties the period effect to observable macro data. Any practitioner would reach for it first. We give it the same network architecture, data splits and tuning budget as our estimator. The only difference is whether the period slope is kept, pinned by the macro covariates, or removed. On predictive fit, the file reports only that the neural model beats vintage-level and aggregate linear models on pseudo-R². No reported pseudo-R² value was found on Lacuna for the Freddie Mac data, so we will re-estimate it ourselves.

Beating this competitor means one of two outcomes. The first is equal or better certainty-equivalent surplus with a lower spread of transfers across identifying assumptions. The second is a smaller surplus loss, below a pre-registered 10 percent of the gain, in exchange for at least halving that spread.

For the synthetic Stackelberg checks, we first reproduce two published reference points before running any indexing experiment. One is the order-quantity drop from 330 to 269 units reported by Wu et al. as retailer loss aversion rises from 1 to 4. The other is the $R_s < 2R_r$ coordination bound of Cai et al..

Two ablations isolate the key ingredients. The first swaps the identified-only index for full indexing under three different identifying constraints: equal adjacent periods, zero period slope, and macro anchoring. This shows whether the identification choice actually moves money. The second replaces joinpoint-selected knots with knots fixed at known event dates, which separates the value of data-driven break detection from the value of indexing itself.

Why not just…?

Why not just index payments directly to unemployment or house prices? This is the most serious threat, and it may partly win. An index built on public macro series is simple and hard to manipulate. However, it insures only the part of the period shock those series explain, so the supplier keeps the remaining basis risk. And because the macro series has its own trend, using it silently fixes the unidentified slope. The macro-index arm of the ablations measures both effects directly. If it matches identified-only indexing on surplus and transfer stability, we will report that the contribution reduces to a diagnostic of when macro indexing is safe, and not to a better contract.

Why not just use time dummies at known disruption dates? Fixed dummies assume the analyst knows when breaks happened. The research file documents that operational volatility often comes from internal scheduling, not calendar milestones (SupplyGraph), and that responses lead or lag announcements. Our fixed-knot ablation tests this directly. If fixed knots at the 2008–2009 dates do as well on Freddie Mac, we keep them for that dataset and confine the joinpoint claim to the battery setting, where breaks are not dated in advance.

Why not just keep simple contracts and assume risk neutrality? Under risk neutrality our contract provably adds nothing, and we report that as the reference point rather than hide it. The case for indexing rests entirely on risk aversion. Bayesian Analysis of Linear Contracts shows that simple contracts lose their near-optimality exactly when agents demand risk premia. The dose-response experiment shows how much aversion is needed before the extra complexity pays.

Why not just learn the contract end-to-end with reinforcement learning? A learned pricing policy would not separate controllable from uncontrollable risk in a form a counterparty could audit. Online Learning in Supply-Chain Games also shows slow convergence and a 4/3 price of anarchy when learners ignore incentives. We treat this as out of scope rather than as a baseline we expect to beat.

Proposed Experiments

The first decisive check is a dose-response curve on synthetic chains. We vary agent risk tolerance from neutral to strongly averse, and vary the share of liability variance coming from period curvature from 0 to 0.8. We then plot the certainty-equivalent surplus gain of identified-only indexing. The prediction fails if the gain does not rise with both variables, or if it is nonzero under neutrality.

The second check replays Freddie Mac hazards. We treat each origination channel as a supplier whose repurchase liability is computed under the four contracts: unindexed, macro-indexed, fully indexed, and identified-only. Each supplier's index is fitted leave-one-out. We report CVaR at the 95th percentile of supplier liability, mean-variance certainty-equivalent surplus, and the share of each supplier's liability attributable to its own cohort effect. That last measure should stay high, so that underwriting incentives are preserved. We also report the range of transfers across identifying constraints, and the specific 2008–2010 periods where full and identified-only indexing disagree most.

The third check repeats the comparison on the battery warranty chain with injected, undated period breaks. It records whether joinpoint selection recovers the breaks under autocorrelated errors, and how much surplus is lost when knots are misplaced.

The idea is supported if three things hold at once. Identified-only indexing captures most of full indexing's premium reduction. It shows markedly smaller transfer swings across identifying assumptions. And it leaves cohort risk with the supplier. If macro indexing matches it on all three, the honest conclusion is that the identification problem does not matter for contract design in these data, and that would be worth knowing.