Do Hydrological LSTMs Keep Telling the Truth About Storage as Lead Time Grows?

Recurrent networks for river forecasting keep a running internal memory. Earlier work found that parts of this memory line up with physical water stores such as snowpack and soil moisture, and hydrologists increasingly read these cells as evidence that the model "understands" the catchment. That evidence comes entirely from one-day-ahead predictions, and it is correlational: a cell tracks a storage estimate, but no one has checked whether the forecast actually depends on that cell the way it depends on real storage. This proposal asks a sharper question, borrowed from clinical interpretability, where an explanation is only useful if it is faithful for the specific case in front of the user. As an LSTM forecasts 1, 3, 7 and 10 days ahead, does its internal storage variable still drive the forecast in the physically correct direction and by roughly the right amount? Or does it keep correlating with storage while the forecast quietly stops depending on it?

The core idea is to probe the network's memory for a storage direction, then push the memory along that direction and check the forecast. Adding water to a catchment should never lower future discharge, and a process model tells us roughly how much it should raise it at each lead time. Measuring this interventional faithfulness at every lead time, basin and season gives two things. The first is a diagnosis of horizon decay: whether accuracy erodes because the internal storage story breaks, or because forcing and readout errors grow around an intact state. The second is a per-forecast trust signal that a river manager could use, much as a clinician needs to know when a model's stated reasoning can be relied on.

Core Claim

We claim that correlational state correspondence and interventional faithfulness come apart as lead time grows. Linear probes from LSTM memory to storage should stay accurate across the horizon. The sign and size of the forecast's response to a storage edit should drift away from physical expectations at longer leads, and fastest in snowmelt and flash-response regimes. We further claim that a per-forecast faithfulness score predicts forecast error beyond what ensemble spread predicts. If either claim fails, the failure is informative. If faithfulness holds across leads, horizon decay in these models is a forcing and readout problem rather than a state problem, which redirects effort toward the input-uncertainty side documented in the Baihe study.

Grounding in Prior Work

Two papers motivate the question. NeuralHydrology - Interpreting LSTMs in Hydrology trained a small LSTM on CAMELS basins and showed that individual memory cells correlate with Upper and Lower Zone Storage and snow water equivalent simulated by SAC-SMA and Snow-17. It also used Integrated Gradients to show seasonal changes in how far back inputs matter. That analysis covers single-step prediction, two basins, and correlation only. We extend it to many lead times and replace correlation with intervention. A New Custom Deep Learning Model Coupled with a Flood Index for Multi-Step-Ahead Flood Forecasting is the second motivating paper. It ran an ANOVA decomposition on Baihe floods and found input uncertainty rising from 32% to 53% over five lead days while structural uncertainty fell from 32% to 28%. It treated the recurrent units as opaque, so it cannot say whether the internal state stays physically grounded. Our audit answers that question from inside the model. Transformer Based Water Level Prediction in Poyang Lake, China documents sharp error growth past 7 days, with RMSE going from 0.46–0.73 m at 1–7 days to 1.82 m at 30 days, but leaves the cause unexplained. It motivates asking why decay accelerates.

Three papers supply the models we audit and set up a tension we can test. MC-LSTM: Mass-Conserving LSTM builds conservation into the cell and finds an accumulator that tracks snow without supervision. It improves peak bias but slightly lowers aggregate efficiency, and it was evaluated only at single steps. Update hydrological states or meteorological forcings? introduces δHBV, a differentiable conceptual model whose states are physical by construction. With data assimilation it matches an unconstrained LSTM at a 1-day lead, but it was also studied only at that lead. WEATHER-5K shows unconstrained forecasters collapsing past 72 hours relative to physics models. It is atmospheric, not catchment, and it did not look inside the models. Together these raise an open question: does a conservation constraint buy faithfulness that survives the horizon, even when it costs efficiency? Our audit measures that directly.

The clinical literature supplies the evaluation standard rather than a method to copy. Safe and Interpretable Estimation of Optimal Treatment Regimes replaces raw telemetry with mechanistic drug-concentration states so that recommendations rest on biologically meaningful quantities. We ask the analogous question of whether a network's learned storage quantity really carries the causal load. PECLIDES Neuro converts ensembles into auditable rules that clinicians can inspect case by case. This supports our emphasis on a per-forecast audit rather than a global importance score. The clinical literature's documented weakness, that post-hoc attributions can be unstable and unfaithful to a sequence model's reasoning, is the specific failure we test for in hydrology. We do not adopt reluctant additive modeling from RAITR. It is designed for static treatment rules, and grafting it onto recurrent gates would be a separate project.

Research Program

The forecasting setup follows an operational pattern on CAMELS. At issue time $t$, the model has seen observed discharge and forcings up to $t$ and receives forcings for $t+1$ through $t+k$. It predicts discharge at every lead $k$ from 1 to 10 days. Forcings for future days come in two versions. The first uses observed forcings, which removes forcing error entirely. The second adds lead-dependent noise and bias to those forcings, calibrated to mimic growing meteorological forecast error. Comparing the two isolates how much decay the model's own dynamics create, which the ANOVA approach in the Baihe study could only estimate from outside.

For each trained model and each lead $k$, we fit linear probes from the memory vector $c_t$ at issue time to reference storages. These are SAC-SMA/Snow-17 states as in the NeuralHydrology study, and separately δHBV's states, so that conclusions do not hinge on one conceptual model. The probe for storage $S$ gives a direction $w_S$. The intervention adds $\delta \, w_S / \lVert w_S \rVert^2$ to $c_t$, which in probe units corresponds to adding $\delta$ mm of storage. We then roll the forecast forward and record the change in discharge at each lead. Faithfulness at lead $k$ is summarized by two quantities. The first is the sign agreement rate: the fraction of forecasts where added storage raises discharge. The second is a response ratio,

$$ \rho_k = \frac{\Delta \hat{Q}^{\text{LSTM}}_{t+k}}{\Delta Q^{\text{ref}}_{t+k}}, $$

where $\Delta \hat{Q}^{\text{LSTM}}_{t+k}$ is the network's discharge change at lead $k$ after the edit. $\Delta Q^{\text{ref}}_{t+k}$ is the reference model's change after receiving the same storage increment in the same basin state. A faithful state has sign agreement near one and $\rho_k$ near one. We do not treat the conceptual model as ground truth for magnitude, so $\rho_k$ is read as agreement with a physical reference. The sign test needs no reference model at all. To check that edits stay on the data manifold, we compare the edited states against the natural distribution of states in the same basin and season, and we report the results at several values of $\delta$.

We audit a standard LSTM, an Entity-Aware LSTM, MC-LSTM (whose accumulator can be edited directly as well as through a probe), and δHBV (where edits are exact). This yields curves of probe accuracy and faithfulness against lead time, stratified by snow fraction, aridity and flow regime. It also yields a per-forecast faithfulness score, computed cheaply from one small edit, that we test as a predictor of that forecast's error.

Baselines

The naive reference point is persistence: the last observed discharge carried forward to every lead. It sets the floor that horizon decay is measured against. A second reference is the same LSTM without discharge input, matching the EA-LSTM setup in The proper care and feeding of CAMELS, which reports a median NSE of 0.74 across 531 basins at one step. As lead grows, a discharge-informed model should converge toward this no-discharge model, which gives a meaningful asymptote.

The strongest existing method on this question is δHBV with data assimilation from Update hydrological states or meteorological forcings?. It reports a 1-day-lead median NSE of 0.82 on CAMELS, up from 0.75, matching LSTM with data assimilation, with states that are physical by construction. It is the strongest competitor because it offers physically interpretable forecasts without any audit. We found no reported result on Lacuna for δHBV or any LSTM at multi-day CAMELS leads, so we run all models ourselves. For the peak-flow comparison, the relevant reported figure is MC-LSTM's top-2% peak flow bias of -14.7 against -15.7 for the LSTM ensemble in MC-LSTM.

For fairness, all models use the same 531 basins, the same train and test periods, the same forcing versions, and the same hyperparameter search budget in GPU-hours. Assimilation models get the same observation window, and results are ensembled over the same number of seeds. "Beating" δHBV is not the goal on accuracy. The proposal succeeds if its per-forecast faithfulness score predicts forecast error better than seed-ensemble spread, measured as rank correlation with absolute error and as area under the curve for flagging the worst decile of forecasts, and if this holds at leads of 5 days and beyond.

The key ablation removes the intervention. We use probe accuracy alone as the per-forecast trust signal and compare it with the interventional score. If correlation predicts error just as well, the interventional ingredient adds nothing.

Why not just…?

Why not just track probe correlations at each lead time, as the NeuralHydrology study did at one step? That is cheaper and would be our first plot. But correlation cannot tell whether the forecast uses the storage variable, and the clinical faithfulness concern is exactly that readable internals can be causally idle. The ablation above settles this. If probe accuracy stays flat across leads while sign agreement and $\rho_k$ drift, the two measures come apart as we claim. If they move together and predict error equally well, we will report that the simpler measure suffices.

Why not just use δHBV or MC-LSTM, whose states are physical by design? This is a real threat. If δHBV matches LSTM accuracy at every lead, a practitioner might skip auditing altogether. Two things keep the question open. First, δHBV was only shown to match at a 1-day lead, and whether its advantage holds, grows or shrinks at 5–10 days is unmeasured. Second, MC-LSTM trades aggregate efficiency for conservation. Because both are audited in the same way, the experiments directly show whether constraints buy lead-stable faithfulness and at what accuracy cost. If constrained models dominate on both, that becomes the headline finding.

Why not just compute Integrated Gradients or SHAP per lead time? Input attributions describe which past days of rain matter, not whether the internal storage drives the forecast. The clinical literature warns that such attributions can be unstable for sequence models. We include Integrated Gradients per lead as a comparator trust signal. We test its own faithfulness with a deletion check: remove the most-attributed input days and measure the forecast change. We then report whether it predicts error as well as the state intervention does.

Proposed Experiments

The first experiment produces decay curves. We plot NSE, peak flow bias, probe accuracy, sign agreement and $\rho_k$ against lead for all four models under both forcing versions, stratified by snow fraction. Running the observed-forcing version makes the decomposition concrete: any faithfulness decay that appears with perfect forcings belongs to the model's dynamics, not to the inputs.

The second experiment tests trust prediction. On held-out years, we correlate the per-forecast faithfulness score with absolute error at each lead and compare it against ensemble spread, probe accuracy and the Integrated Gradients signal.

The third experiment checks whether the findings transfer to a different system. We repeat the audit on the Poyang Lake data from the Transformer water level study, where decay accelerates past 7 days. We use LSTM models and the boundary Yangtze level as an input whose edit has an unambiguous expected sign, and ask whether faithfulness breaks near the same lead where error accelerates.

The decisive check is simple. If sign agreement for the standard LSTM falls clearly below that of δHBV at leads of 5 days or more while probe accuracy stays high, and the faithfulness score flags high-error forecasts better than ensemble spread, the claim stands. If none of these holds, we conclude that learned storage states in hydrological LSTMs are trustworthy across the horizon. That would also be worth knowing, and it would support the Baihe study's attribution of long-lead error to inputs.