When an automated system rejects someone, it often adds a counterfactual explanation: a statement of what they could change to get a different decision. A strong recent idea checks that advice against real life. It compares the suggested change with changes that people have actually made in longitudinal data, meaning records of the same individuals observed over time. Advice that resembles no observed change is ruled implausible. This fixes a real problem. Standard distance-minimizing explanations told 74% of MIMIC-III intensive-care patients to make changes never seen in the data (Using Longitudinal Data for Plausible Counterfactual Explanations). The same check has a hidden weakness, though. The historical record of changes is treated as one timeless pool. When the world drifts, through wage growth, inflation, or a patient's physiology evolving over a hospital stay, the changes that were typical five years or five hours ago are no longer the ones a person can realistically make today.
We propose to treat the set of plausible changes as a quantity that drifts over time and must be estimated for the present. The method has two steps. First, it estimates how the typical change, meaning its center and spread, has drifted over calendar time, using an estimator that adapts to the drift without knowing in advance how fast things change. It then shifts each old observed change into today's conditions before using it as evidence of plausibility. Second, it sets the plausibility threshold using the true number of effectively independent observations. Repeated changes by the same person are correlated, so a pool of ten thousand transitions may carry the information of far fewer. In two sentences: longitudinal grounding of recourse silently assumes the world is not changing, and we correct that assumption by transporting historical changes to the present and calibrating plausibility with dependence-aware sample sizes. The test is whether the resulting advice matches the changes people actually make after the advice is issued, and whether it still flips the decision when carried out.
Core Claim
Three bodies of work each hold one piece of this problem. First, work on the actionability gap shows that feasible advice must stay on the data distribution. FACE builds paths through dense regions of a nearest-neighbour graph (FACE: Feasible and Actionable Counterfactual Explanations). DACE penalizes Mahalanobis distance and outlier scores inside an exact solver (DACE: Distribution-Aware Counterfactual Explanation by Mixed-Integer Linear Optimization). Asemota and Hooker ground plausibility in observed transitions. All three assume the reference distribution is fixed.
Second, work on recourse over time shows that the fixed-world assumption is costly. Time Can Invalidate Algorithmic Recourse reports that sub-population causal recourse keeps only about 39% validity after 100 steps of synthetic drift, while its forecasting-based T-SAR keeps about 72%. Setting the Right Expectations: Algorithmic Recourse Over Time documents "moving goalposts" in competitive settings. These papers address whether advice still flips the decision later, which we call validity. They do not ask whether the advice remains a realistic change, which we call plausibility.
Third, work on non-stationary estimation supplies the tool. Adaptive Estimation and Learning under Temporal Distribution Shift shows that denoising a drifting sequence by soft-thresholding its wavelet coefficients achieves minimax-optimal error for estimating the current value, with no drift rate supplied. Soft-thresholding here means shrinking small wavelet coefficients toward zero. That guarantee assumes independent noise, which longitudinal panels violate.
The new claim is that plausibility under drift is a present-tense estimation problem. Solving it requires an adaptive drift estimator and a dependence-corrected calibration step, and doing so recovers plausibility on changes observed after the advice is issued. Neither pooled nor recent-only transition banks achieve this. Asemota and Hooker pool all transitions. T-SAR forecasts the individual's future state but does not ground actions in observed changes. RecAD models autoregressive dynamics but assumes they are stationary (Algorithmic Recourse in Abnormal Multivariate Time Series). None of these estimates how the space of realistic actions itself moves.
Research Program
Each historical transition is a change vector $\Delta x_i$ observed for one individual starting at time $t_i$. We treat the per-feature mean $m_j(t)$ and scale $s_j(t)$ of transitions as drifting sequences. Here $j$ indexes features, and the estimates can optionally be computed within strata such as age band or admission type. We estimate each sequence with wavelet soft-thresholding (Haar and Daubechies DB8, following the adaptive-estimation paper) applied to the per-time-step empirical averages. Each old transition is then moved into the present regime at the current time $t_0$:
$$\tilde{\Delta x}_{ij} = \frac{\Delta x_{ij} - \hat m_j(t_i)}{\hat s_j(t_i)}\,\hat s_j(t_0) + \hat m_j(t_0),$$
where $\hat m_j$ and $\hat s_j$ are the wavelet estimates. This step resembles reversible instance normalization (Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift), which normalizes a model's input window and restores its statistics at the output. Our version differs in two ways: the statistics are estimated across calendar time rather than from a single look-back window, and they are applied to the reference set of feasible actions rather than to a forecaster's input.
The transported pool replaces the raw pool in the longitudinal distance of Asemota and Hooker. That distance measures how far a proposed change $e - x$ lies from its $s$ nearest observed transitions, scaled by the median absolute deviation. We use it both as a post-hoc ranking of candidate explanations and as a penalty inside their genetic search, so the search procedure stays identical and only the reference set changes.
The second component is the threshold. We accept a proposal only if its longitudinal distance falls below a quantile $q_\alpha$ of the distances between held-out transitions and the rest of the pool. Transitions by the same person at nearby times are correlated. A naive empirical quantile is then overconfident, because its error shrinks with the raw count rather than with the effective sample size. We therefore compute the quantile on blocks of transitions separated by buffer gaps. This is the decoupling device behind Structural Risk Minimization for Nonparametric Time Series Prediction, Stability Bound for Stationary Phi-mixing and Beta-mixing Processes, and the buffer-zone argument of Sharper Bounds for Uniformly Stable Algorithms with Stationary Mixing Process. The gap length is chosen from the empirical autocorrelation of each individual's transitions. With $N$ transitions, blocks of length $a$, and gaps of length $b$, the effective size is roughly $N/(a+b)$, and quantile error should scale as $\sqrt{(a+b)/N}$ rather than $1/\sqrt{N}$. This scaling follows from the standard blocking reduction to approximately independent blocks.
A secondary theoretical aim addresses the open problem stated in the adaptive-estimation paper, which is how the wavelet estimator behaves when its noise is autocorrelated. We conjecture that under β-mixing noise the pointwise rate degrades only through the effective sample size. We will attempt a proof by blocking the wavelet coefficients. The empirical program does not depend on the proof succeeding.
Validity is handled separately and conservatively. We check whether the advice flips the decision of the model deployed at execution time $t_0+\tau$. Where the benchmark provides it, we also combine our plausibility layer with T-SAR's forecast of the user's state. This separation lets us attribute gains correctly: the transport step targets plausibility, and forecasting targets validity.
A known limit is that the transport step models only first- and second-order drift in each feature. This mirrors the limitation that Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting and RevIN share, since both correct only mean and variance. We test this limit directly with a regime in which the correlation between features changes.
Baselines
The naive reference point is distance-minimizing counterfactual search with hand-set immutable features. It runs on the same classifier, with the same candidate budget and the same genetic search. Asemota and Hooker report that such methods proposed impossible changes for 73–84% of Adult individuals and changes never seen in the data for 74% of MIMIC-III patients. We expect it to fail on plausibility and include it to set the scale.
The closest method is pooled longitudinal grounding exactly as in Using Longitudinal Data for Plausible Counterfactual Explanations. Its reported MIMIC-III result is that validity falls from all 229 patients when any change is allowed to 120 patients when changes are restricted to realistic transitions. We rerun it with identical $s$, identical median-absolute-deviation scaling, and identical search generations. The only difference from our method is the reference pool. Beating it means lower longitudinal distance to transitions observed after $t_0$, at equal or higher execution-time validity.
The strongest competitor on validity under drift is T-SAR from Time Can Invalidate Algorithmic Recourse, which reports about 72% validity after 100 steps against about 39% for sub-population causal recourse on its synthetic Loan trend benchmark. We run it on that benchmark with its own forecaster and structural model, and we report its plausibility score alongside ours. Beating it means matching its validity within noise while clearly improving plausibility on future transitions. We do not claim to improve on its validity alone.
Three further baselines and ablations isolate our ingredients. The first is a recent-only pool, which uses transitions from a sliding window of the last $w$ steps and tunes $w$ on validation data with the same tuning budget we give the wavelet threshold. The second is an exponentially weighted transport, which replaces the wavelet estimate of $\hat m_j$ and $\hat s_j$ with an exponential moving average whose decay is tuned the same way. The third is our full method with a naive empirical threshold in place of the blocked one. The first two test whether adaptive drift estimation matters. The third tests whether the dependence correction matters.
Why not just…?
Why not just use only recent transitions? A sliding window is the obvious fix, and on smooth drift it may be adequate. The adaptive-estimation paper argues that sliding windows face an unavoidable bias–variance trade-off, and it reports that wavelet estimators beat an adaptive-window method on Dubai housing data, particularly when validation data was scarce. It gives no exact error table in the material we have. Recourse makes the variance side worse. A short window throws away most transitions, and Asemota and Hooker already show that restricting the transition pool costs validity. The recent-only baseline with a tuned window settles this empirically. We expect it to lose most under abrupt regime changes and when data are scarce.
Why not just normalize features with an exponential moving average and skip wavelets? This is the real threat. If drift in transition statistics is slow and smooth, a tuned exponential average may match the wavelet estimator. The core contribution, transporting the reference pool to the present, would survive that result, but the case for wavelets would not. We will report the gap across drift types: smooth trend, abrupt break, and mixed. If the exponential average matches everywhere, we will say so and recommend it.
Why not just forecast the future state as T-SAR does and ignore plausibility? T-SAR answers whether an action will still work at $t+\tau$. It does not check whether the action resembles anything people currently do. Its structural causal model can still propose large moves in slowly changing root variables, a failure documented in The Importance of Time in Causal Algorithmic Recourse. The two goals are complementary, which is why we also report the combination of T-SAR and our plausibility layer.
Why not just ignore dependence when setting the threshold? On short, weakly correlated panels this may cost nothing. The naive-threshold ablation measures the empirical coverage of the plausibility threshold on held-out future transitions. If naive coverage already meets the target, the blocking step is unnecessary and we will drop it.
Proposed Experiments
We use four settings. The first is the synthetic Loan and additive-noise trend benchmark from T-SAR, extended to a panel. Individuals evolve for 100 steps under a transition kernel whose mean and scale drift, with one variant that adds an abrupt break and another that changes the correlation between features. The second is the linear and Lotka–Volterra systems from RecAD, a predator–prey simulation, with autoregressive coefficients made to drift over time. The third is MIMIC-III, using hourly transitions indexed by hour of stay. Here the drift is along the stay clock, and we extend the Hour 1 to Hour 4 setup of Asemota and Hooker to later windows. The fourth is Adult with simulated aging, which gives comparability to prior numbers.
The decisive metric is future plausibility, defined as the longitudinal distance of each recommended change to transitions actually observed in the window $[t_0, t_0+\tau]$. This held-out window is never used for estimation, so a method cannot score well by fitting the past. We also report execution-time validity, the rate of immutable-feature changes, the empirical coverage of the plausibility threshold, and wall-clock time per instance.
The central check is whether transport lowers future-plausibility distance relative to both the pooled and the recent-only pools, with the gap growing as drift magnitude increases. This should hold on the synthetic panel, where drift is known, and on MIMIC-III, where it is not. On the no-drift control, where the kernel is stationary, pooled grounding should be as good as ours. If transport helps even there, something other than drift correction is driving the result, and we must find it before claiming the effect. On the correlation-change regime we expect transport to help less, and we will report how much less. On coverage, the blocked threshold should hit its target where the naive threshold undercovers in proportion to the measured autocorrelation. If none of these patterns appear, the claim that historical plausibility goes stale is not supported, and the proposal fails honestly.
References
- Using Longitudinal Data for Plausible Counterfactual Explanations
- FACE: Feasible and Actionable Counterfactual Explanations
- DACE: Distribution-Aware Counterfactual Explanation by Mixed-Integer Linear Optimization
- Time Can Invalidate Algorithmic Recourse
- Setting the Right Expectations: Algorithmic Recourse Over Time
- Adaptive Estimation and Learning under Temporal Distribution Shift
- Algorithmic Recourse in Abnormal Multivariate Time Series
- Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift
- Structural Risk Minimization for Nonparametric Time Series Prediction
- Stability Bound for Stationary Phi-mixing and Beta-mixing Processes
- Sharper Bounds for Uniformly Stable Algorithms with Stationary Mixing Process
- Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting
- The Importance of Time in Causal Algorithmic Recourse