Partially Shared Donor Weights for Multi-Metric Synthetic Control

Synthetic control estimates what would have happened to a treated unit, such as a state that passed a tobacco tax, by building a weighted average of untreated "donor" units whose pre-intervention history matches the treated unit's, then projecting that average forward. Its weak spot is short histories. With only a handful of pre-intervention periods, the weights overfit noise. Multi-metric synthetic control (mRSC) addresses this by matching donors on several outcome series at once, for example sales in many departments rather than one. It proves that error falls like $1/\sqrt{K T_0}$, where $K$ is the number of metrics and $T_0$ the number of pre-intervention periods. This is multi-task learning in the plainest sense: several prediction tasks share one set of parameters, here the donor weights.

The gain rests on an all-or-nothing assumption. Every metric must share the same unit-level latent structure, so that one weight vector reconstructs the treated unit on every metric. Real panels rarely behave that way. A store may resemble one donor set in grocery and another in electronics. This proposal replaces "one shared weight vector" with per-metric weight vectors that are pulled toward a common center by an amount learned from the data. The method then borrows strength from auxiliary metrics only to the extent that they agree with the target metric.

Core Claim

The real failure mode of pooling metrics in synthetic control is a mismatch in how each unit loads on the latent factors, not a conflict between gradients. The fix is to shrink each metric's weights toward one another, not to re-weight a single concatenated factorization. For metrics $k = 1, \dots, K$, with $k=1$ the target outcome, let $y^{(k)}_1 \in \mathbb{R}^{T_0}$ be the treated unit's pre-period series. Let $\tilde Y^{(k)}_0$ be the denoised donor matrix for that metric, obtained by singular value thresholding as in Robust Synthetic Control. Let $s_k$ be a per-metric noise scale and $\Delta$ the simplex of non-negative weights summing to one. We solve

$$\min_{w_1,\dots,w_K \in \Delta,\ \bar w \in \Delta}\ \sum_{k=1}^K \frac{1}{s_k^2}\,\big\| y^{(k)}_1 - \tilde Y^{(k)}_0 w_k \big\|^2 \;+\; \sum_{k=1}^K \gamma_k \,\| w_k - \bar w \|^2 .$$

Only $w_1$ produces the counterfactual. Each $\gamma_k$ is a "trust" level for metric $k$, and the two extremes are familiar methods. When every $\gamma_k$ goes to infinity, all metrics share one weight vector, which is essentially mRSC with scale normalization. When every $\gamma_k$ is zero, the target is fit alone, which is single-metric robust synthetic control. The intermediate regime is new. The trust levels are set by empirical Bayes: $\gamma_k$ is proportional to the inverse of the estimated dispersion of $w_k$ around $\bar w$, with one global multiplier chosen by rolling-origin validation on the target's final pre-period block.

The expected behavior follows from a bias–variance argument built on mRSC's rate. If the treated unit's loadings on the target metric differ from the shared loadings by an amount $\delta$, full pooling incurs a bias that grows with $\delta$ and does not shrink with $K$. No pooling pays variance on the order of $1/\sqrt{T_0}$. Partial pooling should track the better of the two and beat both when $\delta$ is moderate. The estimated dispersions also give a continuous transfer diagnostic, in place of mRSC's binary check of whether adding a metric raises the rank.

How This Relates to Prior Work

Two papers motivate the problem. mRSC flattens a units × time × metrics tensor and runs one hard-thresholded SVD, which shows the multi-metric gain. Its only protection against unrelated metrics is a rank test that drops a metric outright. We keep its denoising step and replace its single weight vector with partially shared ones. Mixed Synthetic Nearest Neighbors proves that donor weights are invariant across treatment levels when latent row factors are shared, and uses that invariance to pool data across treatments. Its authors acknowledge that invariance breaks under regime shifts. Our shrinkage is a relaxation of exactly that invariance, applied across metrics instead of treatments. The same estimator could replace the hard invariance in that paper, which we treat as a follow-on rather than the main study.

Two other lines constrain the design. RotoGrad fights negative transfer in neural multi-task learning by rescaling gradients and rotating each task's view of a shared representation. Our setting is a small convex problem with no gradient conflict to homogenize, so we borrow only its spirit: each task gets its own view of shared structure, here a deviation from $\bar w$. Synthetic Controls as Balancing Scores shows that simplex-constrained weights act as balancing scores under a factor model. Keeping every $w_k$ on the simplex preserves interpolation and keeps that argument applicable to the target's own factor model. Gradient-based alternatives that leave the simplex would lose this property. Time-Aware Synthetic Control and ClusterSC address orthogonal weaknesses, temporal dynamics and donor selection. ClusterSC's donor clustering can be composed with our estimator, and we test that combination.

Research Program

The work has three parts.

The first is the estimator. It is a convex program solved by alternating minimization: simplex projections for each $w_k$ and a closed-form update for $\bar w$. We pair it with the empirical-Bayes rule for $\gamma_k$ and placebo-based inference, meaning permutation over donors treated as pseudo-treated units.

The second is analysis. We will make the bias–variance argument rigorous under a factor model in which each metric's loadings split into a shared part and a private part. The target is a finite-sample bound on target-metric post-period error that recovers mRSC's rate when private parts vanish and RSC's rate when shared parts vanish. If the bound cannot be made clean, we will report the analysis as a heuristic and rely on the semi-synthetic study.

The third is empirical. The central study is a controlled semi-synthetic experiment that varies the private share of loadings. It is followed by placebo studies on real panels.

Baselines

The naive reference is single-metric robust synthetic control (Robust Synthetic Control) run on the target metric alone, with the same singular value threshold rule and ridge tuning. It corresponds to $\gamma = 0$ and bounds what ignoring auxiliary metrics costs.

The strongest existing competitor is mRSC. It is the only method in the file that pools metrics for counterfactual estimation, and on Walmart with $T_0$ at 10% of the timeline it reports outperforming RSC by "5–7x in terms of test prediction error." On cricket it reports a median mean absolute percentage error of 2.7% for 5-over-ahead forecasts. We run it from the same denoised inputs, with the same validation budget for its threshold, and with its rank diagnostic enabled. Beating it means lower median placebo post-period error on the target metric across pseudo-treated units, with a paired sign test over units. The paired comparison should hold on Walmart, and the cricket error should not exceed the reported 2.7%. A tie on panels where metrics are genuinely homogeneous would be an acceptable outcome. A loss there would not.

Two ablations isolate the key ingredient. The first is mRSC run after z-scoring each metric, which isolates whether scale normalization alone explains any gain. The second is mRSC with greedy forward selection of metrics by the same validation block, which isolates continuous shrinkage from hard selection. Finally, ClusterSC is run alone and composed with our estimator on the housing price panel, where its paper reports lower median MSE than full-donor synthetic control but no exact value. No reported numerical result was found on Lacuna for that panel.

Why not just…?

Why not just standardize each metric and run mRSC? Standardization fixes differences in scale, which is the problem the research file's Gap 1 emphasizes. It cannot fix a treated unit whose donor mix differs across metrics, because one weight vector is still forced on every metric. The semi-synthetic study tests this directly. At zero private share, standardized mRSC should match our method. As the private share grows, our account predicts that standardized mRSC degrades while ours does not. If standardized mRSC keeps pace, the idea is wrong and we will say so.

Why not just drop bad metrics with mRSC's rank diagnostic or forward selection? This is the most serious competitor. When metrics are either fully shared or fully unrelated, hard selection should be as good as shrinkage. Shrinkage can only win when metrics are partially aligned, which is what the store-by-department structure of Walmart suggests. The forward-selection ablation, with an identical validation budget, settles the question. We report results stratified by estimated dispersion so it is visible where each approach wins.

Why not just fine-tune a multi-task time-series model such as UniTS on the donors? UniTS jointly trains one 3.4M-parameter transformer across 38 datasets and excels in 27 of them. It is purely predictive. It produces no donor weights, no interpolation guarantee, and no balancing-score interpretation, and with $T_0$ of a few dozen points per unit it has little to fine-tune on. We include it as a forecasting reference on Walmart, scored on the same placebo error. A win for UniTS there would show that forecasting accuracy and causal-estimator properties diverge. It would not undercut a weight-based method.

Proposed Experiments

The decisive experiment is semi-synthetic. We generate panels from a rank-$r$ factor model with $K$ metrics, in which each unit's loadings mix a shared component with a metric-specific component at a controlled private share between 0 and 1. We sweep $T_0$ from very short to long and add a known effect after $T_0$. Metrics are bias, root mean squared error of the target-metric effect estimate, and placebo-test coverage. The predicted signature has three parts. Our estimator should match mRSC at zero private share and match single-metric RSC at full private share. It should beat both in between, with the largest margin at short $T_0$. Fitted dispersions should track the true private share more closely than mRSC's rank diagnostic does.

On real data, we use the Walmart panel of 45 stores and 81 departments. We restrict $T_0$ to 10% and 25% of the timeline, treat each store–department in turn as pseudo-treated with no true effect, and report median post-period error against mRSC's reported regime. We repeat the procedure on the cricket trajectories, with runs and wickets as two metrics and runs as the target, at 5-over-ahead horizons. On the 400-metro, 40-quarter housing price panel we pair metros' price series with lagged neighbors' series as auxiliary metrics, run the ClusterSC composition, and report median MSE. California Proposition 99 serves as a sanity check that the estimator reproduces the accepted counterfactual shape when only a few auxiliary series are available.

Two checks would falsify the proposal. The first is if partial pooling fails to beat forward selection anywhere in the semi-synthetic sweep. The second is if, on Walmart, the fitted $\gamma_k$ collapse to the two extremes, since that would mean real metrics are either fully shared or fully unrelated and hard selection is enough.