Persistence Shapley Weights for Composite Indicators

Composite indices such as "urban flood resilience," "ecological vulnerability," or "green development quality" are built by giving each of dozens of measured indicators a weight and adding them up. The most common objective recipe, the Entropy Weight Method (EWM), gives an indicator more weight the more its values spread out across the evaluated regions or firms. That rule has two failures that anyone who has built such an index will recognize. Two near-duplicate indicators both get high weight, so the shared quantity is counted twice. A column of pure measurement noise spreads out a great deal, so it can receive one of the largest weights in the index. This proposal fixes both failures with one construction. Treat the indicators as players in a cooperative game. Let the value of any subset of indicators be how much it tells you about the system's next observed state. Then take each indicator's Shapley value, its average added value across all possible subsets, as its weight.

The intuition fits in two sentences. Noise does not persist over time, so it adds nothing to any subset's ability to anticipate the next period, and the Shapley "dummy" property sends its weight to zero. A duplicate adds nothing beyond its twin, so the Shapley average splits the shared credit between them instead of counting it twice. The method needs no outcome label and no expert judgments, which keeps the property that made entropy weighting popular. It replaces "spread" with "persistent, non-redundant information" as the definition of a useful indicator.

Core Claim

Most of the evaluation settings where EWM is used are panels: 16 cities over 23 years in the Chengdu–Chongqing waterlogging study, 21 indicators over 2000–2020 in the Qilian cryosphere study, 27 provinces over 2006–2015 in the logistics FDI study, and 12 indicators over 2011–2020 in the Anhui study. In panels, the future of the data itself can serve as the target. For a subset $S$ of the $m$ indicators, define

$$v(S) = I\big(\tilde X_S^{(t)};\ \tilde X_N^{(t+1)}\big),$$

where $\tilde X_S^{(t)}$ is the vector of indicators in $S$ for an entity in year $t$ after removing each year's cross-entity mean, $\tilde X_N^{(t+1)}$ is the full indicator vector for the same entity a year later, and $I$ is mutual information, the reduction in uncertainty about one quantity from knowing another. The weight of indicator $j$ is its normalized Shapley value, $w_j = \phi_j(v)/v(N)$. By the Shapley efficiency property, these weights sum to one exactly, with no ad hoc renormalization.

Removing the yearly mean matters. Without it, every indicator that trends upward over the period, which describes most development indicators, would look persistent only because of the common time trend. With it, $v$ measures how well a subset tracks stable differences between entities, and those differences are what a ranking index is supposed to capture.

The new element is the choice of characteristic function, the rule that assigns a value to each subset of players. Shapley values over mutual-information games already exist. Covert et al. use conditional entropy of a supervised label $Y$ to explain predictive uncertainty, and prove that a feature gets zero credit exactly when $Y$ is conditionally independent of it given the rest. We reuse that mechanism and that zero-credit result. We replace the external label with the system's own next state, which turns an explanation tool for a trained model into a label-free weighting rule for an evaluation matrix. Their theorem then gives the property EWM lacks: an indicator gets zero weight exactly when the entity's future state is conditionally independent of it given the other indicators. That is true of independent noise and false of a high-variance but meaningful indicator.

Research Program

Estimator. The first piece of work is an estimator that survives the small sample sizes of real panels. Anhui has only ten annual rows, and Chengdu–Chongqing has about 350 city-year pairs. Nonparametric entropy estimates are hopeless at these sizes. We will use a Gaussian copula estimate: map each indicator to normal scores through its ranks, then compute $v(S) = \tfrac12 \log\big[\det \Sigma_{SS}\,\det \Sigma_{FF} / \det \Sigma_{S\cup F}\big]$. Here $\Sigma$ is a shrinkage-regularized correlation matrix over current indicators $S$ and next-year indicators $F$. Rank transformation also caps the influence of heavy-tailed outliers, which addresses the third known failure of EWM.

Computation. For $m \le 21$, every one of the $2^{21}$ subsets can be evaluated exactly, using incremental Cholesky updates on matrices of at most 42 dimensions. That is minutes of CPU time, so exact Shapley values are available on every benchmark in the file. Permutation sampling is needed only for larger synthetic stress tests.

Robustness to the estimate itself. Shapley values are known to be sensitive to how "absent" players are simulated. The Explanation Game shows that marginal-sampling SHAP gives credit to dummy features when features are correlated. Our game avoids that problem by construction: $v(S)$ is a property of the joint distribution of $X_S$, not of a model queried on artificial inputs. The remaining risk is estimation error, so we will report bootstrap intervals over entities for every weight.

Interpretation and optional structure. We will also characterize the weights analytically in the Gaussian case, covering duplicates, a sum of two indicators alongside its parts, and pure noise. Frye et al. show that uniform weighting over orderings is the maximum-entropy choice under complete ignorance, which is the right default when no structure is known. When an index has a declared hierarchy, such as Pressure–State–Response, we will test restricting orderings to respect it, in the spirit of the Extended Owen value. That paper explains an already-specified hierarchical decision model and does not derive weights from data; for us the hierarchy is an optional constraint, not the main contribution.

Stated limitation. By design, the method down-weights variation that is real but transient, such as a one-year rainfall shock. We treat this as part of what an index of a durable state should do, and we will report which real indicators lose weight for this reason so domain experts can judge it.

Baselines

The naive reference point is the standard four-step EWM, implemented as in the logistics FDI study. We will validate our implementation by reproducing the reported Chengdu–Chongqing resilience trajectory, a mean index rising from 0.158 to 0.525 over 2000–2023 (waterlogging study), if the raw panel can be obtained. The file reports no individual entropy weights or CRITIC weights for any dataset ("no reported result found on Lacuna" for either), so every weight comparison will be computed by us on identical inputs.

The strongest existing objective method is CRITIC, as used in the Qilian study. CRITIC multiplies each indicator's spread by its total linear disagreement with the other indicators, so it is the only standard objective method that already targets redundancy. We will also run a strengthened variant we construct ourselves, reliability-CRITIC. It multiplies CRITIC weights by each indicator's own detrended lag-one autocorrelation, which gives CRITIC a cheap noise filter. Reliability-CRITIC is the real competitor, because it attacks both failures with pairwise statistics and no game theory. A third baseline is first-principal-component and factor-analysis loading weights. A fourth is the sequential pipeline of the tennis momentum paper: EWM-TOPSIS scores, an XGBoost model fit to them, and mean absolute SHAP values used as weights. Its reported test $R^2$ of 0.9777 measures how well XGBoost reproduces the EWM-derived score. It is not a measure of weight quality and cannot be compared directly with our results.

Three ablations isolate the proposal's ingredients. The first keeps the Shapley game but uses $v(S) =$ the Gaussian joint entropy of $X_S$. This removes persistence and should fix duplication but not noise. The second keeps the persistence target but uses leave-one-out marginal contributions instead of Shapley averaging. This should fix noise but zero out duplicates. The third keeps everything but skips the yearly demeaning, which tests the trend confound.

All methods see the same rank-transformed or min–max-normalized matrices and receive no tuning on evaluation targets. The only hyperparameter, the shrinkage level, is chosen by a fixed rule (Ledoit–Wolf). We count as beating the baselines a statistically reliable win, by bootstrap over entities and simulation seeds, on latent-state recovery and on weight placed on injected noise. On clean data with no duplicates and no noise, we require our method to be no worse than CRITIC.

Why not just…?

Why not just use CRITIC, or CRITIC multiplied by autocorrelation? For the cases people usually test, pairwise redundancy plus a reliability factor may genuinely work as well, and we say so. Where it should break is multi-way redundancy. If indicator C is approximately A + B, with each pairwise correlation only moderate, CRITIC sees no strong duplicate and counts all three. The Shapley game sees that C adds little once A and B are present. Pairwise autocorrelation also cannot distinguish a persistent indicator that is redundant given others from one that is unique. The synthetic "sum-of-parts" and "three-way overlap" conditions are designed to settle this. If reliability-CRITIC matches us there too, the honest conclusion is that the game-theoretic layer is unnecessary, and we will report that.

Why not just run PCA or factor analysis first? These methods reward shared variance, so a duplicated indicator inflates the leading component and gets more weight, not less. They also assume a single linear factor. Hard pruning, as in the incomplete soft set method that cuts 12 food-safety attributes to 5, discards partial information entirely. The duplicate-injection experiment measures directly whether loading weights double-count.

Why not just shuffle or drop each indicator and see how much the index changes? Leave-one-out gives zero importance to both members of a redundant pair, the classic failure described in the cooperative-game classification paper. Stochastic importance methods also show instability: only 5 of 50 Diabetes instances reached stable hypotheses for LIME, against 50 of 50 for deterministic SHAP (Weight of Evidence study). Our second ablation is exactly this alternative applied to our own value function.

Why not just keep the EWM → XGBoost → SHAP pipeline? The model is trained to reproduce an index whose weights came from EWM. Its attributions therefore inherit EWM's noise and duplication errors rather than correcting them. Running this pipeline with injected noise columns will show whether the noise survives into the SHAP-derived weights.

Proposed Experiments

The decisive test is synthetic, because ground truth exists only there. We simulate panels from a dynamic latent-factor model with known persistent entity states. Indicators are noisy linear and monotone nonlinear readouts of those states. We vary the number of entities (10 to 400) and years (10 to 25) to match the real datasets. We then inject exact and near duplicates, sum-of-parts indicators, independent high-variance noise, and heavy-tailed outliers.

The primary metric is Spearman correlation between each method's composite index and the true latent ranking. Secondary metrics are the total weight placed on injected noise and the summed weight of a duplicated pair relative to the original indicator. The predicted signature is specific. EWM puts large weight on noise. The joint-entropy ablation splits duplicates but keeps noise. The leave-one-out ablation removes noise but zeroes duplicates. Only the full method does both.

The semi-synthetic test uses the real indicator matrices: Chengdu–Chongqing (19 indicators), Qilian (21), logistics FDI (7) and Anhui (12), where raw data can be obtained. We append duplicate and noise columns and measure how much the entity rankings move, using Kendall's tau between the original and augmented indices. A good weighting method should leave the ranking almost unchanged.

On the unmodified real data, we report two external validity checks. The first is the rank stability of the index between adjacent years. The second is agreement with independent evidence: in Qilian, overlap of our top-weighted indicators with the five drivers that random forest plus SHAP identified, namely vegetation cover, biological abundance, glacial meltwater, precipitation and landscape diversity (Qilian study).

We will consider the proposal refuted if reliability-CRITIC matches it within bootstrap error on the multi-way redundancy conditions. We will also consider it refuted if the copula estimator's weight intervals at Anhui-scale sample sizes are too wide to rank indicators. In that case the method would be restricted to the larger panels, and we would say so.