Allocating Under Gaming: Using Past Policy Changes to Find the Causes That Matter for the Decision

Banks, grant agencies and hospitals often run a two-step pipeline. A model predicts a number for each applicant, such as expected loss, project yield or clinical need, and an optimizer uses those numbers to split a fixed budget. Applicants learn what the model rewards and change their features in response. Some changes are genuine: paying down debt lowers the true loss. Others are gaming: opening extra credit lines lowers a utilization ratio without making anyone safer. Methods that train predictors directly for decision quality assume the true costs do not react to the model. Methods that tell gaming apart from improvement assume someone already knows which features cause the outcome. We propose to drop both assumptions. The core observation is that every time the institution changes its deployed model, it changes applicants' incentives and nothing else in the world. Each redeployment therefore acts as a free intervention on applicant behaviour. A feature that truly causes the outcome keeps a stable relationship with it across these rounds. A gamed proxy's relationship with the outcome breaks exactly in the rounds when it was rewarded.

We turn that signature into a training method for the budget-allocating predictor, and we make the causal search targeted. The method only resolves the causal role of features whose error would change which applicants get funded. It deliberately leaves decision-irrelevant parts of the causal graph unresolved. In two sentences: past deployments give you the interventions needed to tell causes from proxies. You should spend that evidence only on the features that move the allocation, and train the predictor to minimize regret on applicants' post-response true costs.

Where the three literatures stand

The gaming-improvement line supplies the problem and the obstacle. Strategic Classification is Causal Modeling in Disguise casts applicant adaptation as intervention in a structural causal model, meaning a set of equations saying which variables cause which. It proves that designing a rule that incentivizes improvement is as hard as orienting the causal graph. That proof is purely theoretical and assumes a known skeleton. Three later papers build working methods but take the causal roles as given. Learning to Incentivize Improvements from Strategic Agents assumes a known taxonomy of improvable, manipulable and immutable features. Causal Strategic Classification: A Tale of Two Shifts assumes a known causal/correlative partition. Improvement-Focused Causal Recourse assumes a full causal model. The closest work is Disentangling misreporting from genuine adaptation in strategic settings. It separates misreporting from real change because only real change propagates to a downstream outcome. It needs a known descendant variable and a clean, never-manipulated baseline dataset. We need neither: we replace the clean baseline with variation across the institution's own sequence of policies. All of these works end in a scalar accept/reject decision, not a budgeted optimization.

The sparse causal discovery line supplies the mechanism. Causal Discovery in Heterogeneous Environments Under the Sparse Mechanism Shift Hypothesis scores candidate graphs by how few conditional distributions change between pairs of environments. The hypothesis behind it is that real shifts touch only a few mechanisms. Scoring environment pairs rather than pooling them lifts orientation recall toward 1.0, where pooled PC stalls near 0.6. That work needs labelled environments, which is exactly what deployment rounds provide for free. It is also untargeted, spending samples on every edge equally. Identifiable Nonlinear Differentiable Causal Discovery via Independence and Adaptive Group Sparsity constrains our design. It shows that mean-squared-error-based continuous discovery misorients edges under unequal noise, and its independence-based residual test is what we reuse inside each round. BaCaDI shows that unknown intervention targets can be inferred jointly with the graph. In our setting the targets are partly known, because the institution knows which features its policy rewarded.

The predict-then-optimize line supplies the objective. Smart "Predict, then Optimize" introduced regret, the extra true cost of acting on predicted rather than true numbers. It also introduced SPO+, a convex upper bound on regret that can be trained by gradient descent. Melding the Data-Decisions Pipeline shows the accuracy paradox: a decision-trained model had roughly seventeen times worse squared error yet a much better objective (49.18 versus 32.13 on budget allocation). Solver-Free Decision-Focused Learning for Linear Optimization Problems removes the solver from the training loop. All three treat true costs as fixed, which is precisely what strategic applicants violate.

Core Claim

Let rounds $t = 1, \dots, T$ each deploy a predictor $f_{\theta_t}$. Applicants respond, the institution solves $\min_{w \in \mathcal{S}} \hat c^\top w$ over allocations $w$ in feasible set $\mathcal{S}$, and true costs $c$ are observed for funded applicants. Because only the policy changes between rounds, the sparse-shift hypothesis predicts a specific pattern. The only mechanisms that shift are applicant action mechanisms on rewarded features, plus the outcome conditional $P(c \mid x_j, \cdot)$ for any rewarded feature $x_j$ that is not a cause of $c$.

We estimate, for each feature $j$, a shift score $s_j$. It is the pairwise mechanism-shift evidence that the outcome's dependence on $x_j$ broke across rounds. We then train

$$ \min_\theta \; \sum_t \sum_i \ell_{\text{SPO+}}\big(f_\theta(x_i^{(t)}),\, c_i^{(t)}\big) \;+\; \lambda \sum_j \pi_j(\theta)\, s_j\, \|\theta_j\|, $$

The terms are as follows. $\ell_{\text{SPO+}}$ is the regret surrogate on post-response data. $\theta_j$ is the block of weights reading feature $j$. $\pi_j(\theta)$ is a decision-sensitivity weight: the fraction of instances in which perturbing the prediction along feature $j$'s contribution moves the cost vector across the boundary of the current optimal allocation's optimality region. That region can be computed from adjacent vertices, as in the solver-free method. The penalty is large only for features that are both shift-flagged and decision-relevant.

The claimed contribution has three parts. First, deployment history identifies the causal roles that matter, with no clean baseline and no given taxonomy. Second, targeting that identification by decision sensitivity needs fewer rounds and samples than full-graph discovery for the same regret. Third, the institution can choose the next policy to probe the still-ambiguous, decision-relevant features, turning Strategic Classification is Causal Modeling in Disguise's abstract oracle queries into bounded-regret experiments.

The same paper warns that, when costs are complementary, weight on a non-causal proxy can still induce improvement. That is why we penalize rather than delete flagged features: the regret term, computed on true post-response costs, decides whether a proxy earns its weight.

Research Program

The first component is a simulator. It pairs structural causal models over applicant features with applicants who best-respond under a quadratic action cost and can act on causes or proxies. Two relaxations of best response are included: the local-gradient responses of Non-linear Welfare-Aware Strategic Learning, and noisy responses. Downstream, a knapsack allocates loans or grants under a budget, reusing the instances and adjacent-vertex precomputation in ML-KULeuven/Solver-Free-DFL.

The second component is the discovery module. It computes pairwise mechanism-shift scores on the outcome conditional between rounds and uses an independence-of-residuals test rather than squared error, following the constraint from NOTIME-CAM. Only funded applicants reveal costs, so the module must correct for selective labels. We will do this by inverse-propensity weighting with the known allocation rule, and we treat this correction as a first-class risk.

The third component is the regret-weighted training objective above. The fourth is a policy-selection step. It perturbs the next round's weights toward ambiguous features with high $\pi_j$, subject to a cap on expected regret.

The theory target is modest. For linear causal outcome models, we aim to show two things. Features whose rewarded weight changed in at least two rounds are identified as cause or proxy. The number of rounds needed scales with the number of decision-relevant ambiguous features rather than with the total feature count.

Baselines

The naive reference is two-stage retraining. Each round, a model is fit by squared error on the latest post-response data and its predictions are fed to the knapsack solver. This is what most deployments do, and it chases gamed proxies. In the strategic classification analogue, plain risk minimization reached only about 66.7% accuracy against 87.8% for the causal method in Causal Strategic Classification: A Tale of Two Shifts.

The strongest decision-focused competitor is SPO+ retrained every round on post-response data, with LAVA as a faster variant; both run through the same codebase. It is strongest because it optimizes the right objective and adapts each round, but it is blind to causal roles. The reported speed figure is LAVA's knapsack training time of 9.8 seconds against 581 seconds with a solver in the loop, from Solver-Free Decision-Focused Learning for Linear Optimization Problems. For decision regret under strategic applicant response, no reported result was found on Lacuna, so we will establish that number ourselves.

The strongest causal competitor is Constructive Adaptation risk from Learning to Incentivize Improvements from Strategic Agents, given the true feature taxonomy, with its scores fed to the knapsack. On German Credit it reports a 56.10% improvement rate against 33.50% for standard strategic models. We also include a discover-then-optimize pipeline: pooled NOTEARS or PC on all rounds, then SPO+ restricted to the discovered parents.

The upper bound is SPO+ with the oracle causal partition. All methods get identical rounds, applicant populations, feature access, knapsack instances and hyperparameter-search budget. The search budget is a fixed number of configurations, tuned on held-out rounds' post-response regret.

Two ablations isolate the key ingredients. The first sets $\pi_j \equiv 1$, giving untargeted shift-penalized SPO+. The second removes policy selection and learns only from passive rounds.

Our method beats the field if three conditions hold. Its post-response regret must be significantly below per-round SPO+, with seeds as the unit of variance. It must reach within a stated margin of the oracle-partition bound in fewer rounds than the untargeted ablation. Its improvement rate must be at least that of Constructive Adaptation run with the true taxonomy.

Why not just…?

Why not just retrain SPO+ every round and let decision-focused learning absorb the gaming? Per-round retraining sees gamed proxies still correlated with cost in the most recent data. The breakage only shows up after the model rewards the proxy, so the learner lags one round behind the applicants and keeps funding gamers. This is a real threat when applicants respond slowly or weakly. Our decisive comparison varies applicants' responsiveness. If per-round SPO+ matches us at every level, the causal component is unnecessary and we will say so.

Why not just treat manipulation as bounded adversarial noise and train a robust predictor? Robustness suppresses response to every feature change, including genuine improvement. The file's evidence is consistent. Robust strategic models reach a 33.50% improvement rate against 56.10% for the causally aware objective on German Credit in Learning to Incentivize Improvements from Strategic Agents. Causally unaware recourse yields near-zero real improvement in Improvement-Focused Causal Recourse. We will include a robust-SPO+ variant and report both regret and improvement rate for it.

Why not just run PC or NOTEARS once on pooled data and then optimize? Pooling across shifting environments leaves orientations at the equivalence-class level, around 0.6 recall versus near 1.0 for pairwise shift scoring in Causal Discovery in Heterogeneous Environments Under the Sparse Mechanism Shift Hypothesis. On Sachs, NOTEARS reaches a structural Hamming distance of 22 against 14 for the independence-based method in Identifiable Nonlinear Differentiable Causal Discovery via Independence and Adaptive Group Sparsity. The discover-then-optimize baseline tests whether these statistical gaps matter for regret.

Why not just keep only features whose relationship to cost is stable across rounds, with no decision weighting? This is the most serious alternative, because it uses the same invariance signal. It is exactly our $\pi_j \equiv 1$ ablation. Our claim is that decision weighting pays off when many features are ambiguous but few move the allocation. We therefore sweep the ratio of decision-relevant to total features. If the untargeted version matches ours across that sweep, the contribution reduces to "shift-penalized decision-focused learning." That is still new relative to the file, but narrower, and we would report it that way.

Proposed Experiments

The primary testbeds are semi-synthetic, because true post-response costs must be known. We use four. The first builds causal models on the UCI credit card data (30,000 applicants), as in Disentangling misreporting from genuine adaptation and Causal Strategic Classification: A Tale of Two Shifts. The second is German Credit with the taxonomy of Learning to Incentivize Improvements from Strategic Agents. The third is that paper's synthetic DAG agents. The fourth is the 5var-skill and 7var-covid models from Improvement-Focused Causal Recourse. Each is coupled to a capital-budget knapsack from the solver-free codebase. We also run the discovery module on Sachs, a non-strategic dataset with a known graph, only as a sanity check that our shift-scoring implementation is competitive with the reported numbers.

Metrics are four. Post-response regret on true costs. Improvement rate, meaning the share of applicants whose true cost fell. Orientation recall separately on decision-relevant and irrelevant edges. Rounds needed to come within a fixed margin of oracle regret.

Four checks are decisive. First, regret-versus-round curves should show per-round SPO+ oscillating as applicants game, while our method converges. Second, our method's advantage over the untargeted ablation should grow as decision-relevant features become a smaller share of all features. Third, recall on irrelevant edges should stay low for our method without hurting regret; this is the decision-focused analogue of the accuracy paradox, now for causal structure. Fourth, a stress test under hidden confounding between action cost and outcome, where the invariance signal can mislead, should show where the method breaks. Reporting that failure boundary is part of the result.