This work studies context-aware selective transfer learning for multi-well groundwater anomaly forecasting in the Netherlands. Source-domain pretraining can raise average accuracy but is not uniformly reliable: it degrades performance on a substantial fraction of heterogeneous wells. Context-Aware Selective Learning (CASL) uses calibration-trained, constrained contextual blending of frozen no-transfer and fixed-transfer parents. Global Convex Blend uses the same parents with one calibration-fitted weight per seed. The corrected task requires nine consecutive observed monthly inputs and an exact three-calendar-month target, without interpolation. Evaluation covers 457 wells and 13,718 test windows under five predefined seeds, with the four-cell maxsim0.60 source pool frozen using Always-TL validation RMSE alone. Fixed transfer harms 18.38-51.86% of wells. CASL has 0.424-1.084% lower RMSE than No TL, partially rescuing 97.53-99.50% and fully rescuing 23.46-43.78% of fixed-harmed wells. Partial rescue can arise structurally from convex blending and does not establish contextual superiority. Global Convex Blend protects more wells in four seeds, while CASL has slightly lower pooled RMSE in four. Neither dominates: the results support an accuracy-risk trade-off on the fixed split, not rolling deployment validity.
Figures & tables
Figure 1: Pipeline schematic for Context-Aware Selective Learning (CASL).
Parameter
Value
Forecast horizon
3 exact calendar months
Sequence length
9 consecutive observed months
Target splits (train/val/cal/test)
Frozen target-date boundaries
Windows (train/val/cal/test)
41,551 / 10,970 / 6,562 / 13,718
Forecast-origin embargo
None
Source filtering rule
maxsim0.60 (Always-TL validation)
Table 1: Main training configuration used in the corrected experiments. Target splits retain the fixed response-date boundaries; the selected CASL gate is histogram gradient boosting (HGB).
Policy
RMSE (range)
MAE
R2
No TL
2.16856±0.01593
0.93611±0.01351
0.21107±0.01161
[2.15378,2.18791]
Always TL
2.16695±0.02611
0.90227±0.03155
0.21219±0.01894
[2.12938,2.19600]
Global Convex Blend
2.16640±0.01818
0.92863±0.01253
0.21263±0.01323
[2.14830,2.18791]
Table 2: Aggregate held-out metrics across the five predefined seeds on the same 13,718 observations. Entries show mean ± sample SD; bracketed RMSE ranges are observed minima and maxima. Seed variation is conditional on the fixed data and split, not population or temporal uncertainty. Anomaly error units are dimensionless.
Figure 2: Held-out aggregate RMSE, MAE and R2 under the corrected frozen protocol for seeds 11, 23, 42, 67 and 89. Every seed evaluates the same 457 wells and 13,718 test observations. Axis ranges expose close differences; the figure does not establish large or universal superiority. CASL RMSE is lower than No TL in five seeds and lower than Always TL in four. GCB is reported in the main tables, rather than plotted here. Seed 42 is the unchanged reference; no model or seed is selected from these test scores. Observed variation describes training randomness conditional on fixed data/splits, not temporal or population uncertainty.
Outcome
Mean
SD
Median
Range
Always-TL harm (%)
37.94
12.48
40.04
18.38–51.86
Partial rescue (%)
98.48
0.80
98.73
97.53–99.50
Global partial rescue (%)
100.00
0.00
100.00
100.00–100.00
Full rescue (%)
33.68
9.34
32.07
23.46–43.78
Global full rescue (%)
16.62
22.88
0.00
0.00–44.81
Residual CASL harm (%)
26.13
9.23
28.23
10.72–35.45
Table 3: Cross-seed harm/rescue rates. Unqualified rescue, new-harm and net-reduction rows refer to CASL; global-blend rows name the comparator explicitly. Means and sample SDs are unweighted across seeds; SD is in percentage points. Harm percentages use all 457 wells, rescue percentages use the seed-specific fixed-harmed subset, and new harm uses the Always-TL-helped subset. Full rescue is nested within partial rescue; equality with No TL is not strict full rescue. Net reduction is relative to Always TL and accounts for newly introduced harm. Partial rescue is descriptive, not evidence of contextual added value.
Figure 3: Harm and rescue over the five predefined seeds. Always-TL harm and residual CASL harm use all 457 wells; partial/full rescue condition on the seed-specific Always-TL-harmed subset; new harm conditions on wells helped by Always TL. Full rescue is included in partial rescue and the two must not be added. Residual CASL harm includes persistent and new harm. Connected points identify seeds, not temporal change; seed 42 remains the original reference. Rates are retrospective evaluation outcomes and are not model-selection criteria. Partial rescue does not establish contextual-gate value; GCB is compared in the main tables.
Figure 4: Paired effects over the five predefined seeds under the same fixed evaluation. Left: net harmed-well rate reduction, Always-TL harm minus total CASL harm, in percentage points; the latter includes newly introduced harms. Right: relative aggregate RMSE improvements (RMSEparent−RMSECASL)/RMSEparent×100 versus No TL and Always TL. Positive values favor CASL. Every seed has positive net risk reduction and a small RMSE benefit versus No TL; one seed has worse aggregate RMSE than Always TL. These are paired descriptive contrasts against the parents, not evidence of stronger protection than GCB, confidence intervals or temporal robustness.
Figure 5: Distribution of the number of seeds in which a well has each outcome. All 457 wells are retained. Full rescue and persistent harm require Always-TL harm in the same seed; their per-well conditional eligibility counts are retained in the outcome-frequency CSV. Only three wells are harmed by Always TL in all five seeds; no well is fully rescued or persistently harmed in all five. Sixteen distinct wells are newly harmed once each, with no repeated new-harm identity. Recurrence of the CASL-versus-Always-TL outcomes does not imply stable individual identities or superior protection over GCB.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Count
National GLD series with CSV URLs
112,627
After temporal filter (first date ≤ 2005-01-01, last date ≥ 2014-01-01, observations ≥ 500)
2,102
One tube per well after deduplication
1,280
Successful downloads / cached series
1,118
Series with at least 24 monthly anomaly rows in the 1996-07–2016-06 window
754
Series in the original monthly anomaly reference pool
754
Appendix
Table 4: Target-data filtering funnel for the Netherlands modelling panel. The original download and anomaly-panel stages are unchanged; final eligibility and test counts use the corrected calendar protocol.
Candidate
Source cells
Val. RMSE
Val. MAE
Gap to best
Selected
all
512
1.555131
0.766569
0.008591
No
maxsim0.40
230
1.560149
0.771671
0.013608
No
maxsim0.50
69
1.547691
0.753573
0.001150
No
maxsim0.60
4
1.546540
0.750084
0.000000
Yes
Appendix
Table 5: Corrected source-filter selection using only seed-42 Always-TL validation RMSE. The filter is frozen before gate fitting and is not reselected by seed. Gaps use unrounded scores; the leading candidates are numerically near-equivalent.
Policy
RMSE
MAE
R2
Harm n (%)
Partial n (%)
Full n (%)
New n (%)
No TL
2.158150
0.948517
0.218662
—
—
—
—
Always TL
2.154517
0.932157
0.221290
201 (43.98)
—
—
—
Global Convex Blend
2.148304
0.924237
0.225775
124 (27.13)
201 (100.00)
77 (38.31)
0 (0.00)
CASL
2.148994
0.928552
0.225278
122 (26.70)
200 (99.50)
88 (43.78)
9 (3.52)
Site-Hard
2.156212
0.936597
0.220065
—
—
—
—
Appendix
Table 6: Corrected seed-42 aggregate and transfer-risk comparison on 13,718 observations from 457 wells, including GCB. Harm counts/rates use all 457 wells; partial/full rescue use 201 fixed-harmed wells; new harm uses 256 Always-TL-helped wells. Rates are percentages. Equality with No TL is not strict full rescue. Dashes mark reference/inapplicable entries or Site-Hard diagnostics not reported here.
Seed
No TL
Always TL
GCB
CASL
Site-Hard
GCB α
11
2.159294
2.129380
2.159294
2.135883
2.138950
0.000000
23
2.153781
2.170417
2.152840
2.144486
2.168945
0.264697
42
2.158150
2.154517
2.148304
2.148994
2.156212
0.567447
67
2.187909
2.184416
2.187909
2.166285
2.190298
0.000000
89
2.183663
2.196001
2.183663
2.170657
2.195027
0.000000
Appendix
Table 7: Per-seed corrected test RMSE and calibration-only GCB coefficient. Every seed uses identical wells, test windows and frozen source filter. Seed 42 is the retained reference, not a representative realization. Zero α reproduces No TL exactly.
Figure 6: Corrected seed-42 held-out test RMSE, MAE and R2 for matched No TL and Always TL parents, their frozen CASL convex blend and secondary Site-Hard policy. All use 457 wells and 13,718 exact-H3, consecutive-nine-month test windows. Point estimates are close; this panel shows the original four-policy reference comparison only. GCB and the five-seed accuracy–risk comparison are reported in the main comparison tables. WTDA is a dimensionless standardized groundwater-level anomaly.
Figure 7: Illustrative corrected seed-42 diagnostic. Left: strict per-well harm (policy RMSE exceeds No-TL RMSE), denominator 457 wells. CASL residual harm separates 113 persistently harmed and nine newly harmed wells. Right: partial rescue (CASL RMSE below Always-TL RMSE) and full rescue (CASL RMSE below No-TL RMSE), denominator 201 fixed-harmed wells. Full rescue is included in partial rescue; the rescue bars are not disjoint. Partial rescue is descriptive and does not establish contextual added value over global blending.
Figure 8: Illustrative corrected seed-42 diagnostic. Five mutually exclusive outcomes sum to 457 wells: 88 full rescues, 112 partial-but-not-full rescues, one without partial rescue, 247 preserved Always-TL benefits and nine new harms. The 113 persistently harmed wells comprise the 112 partial-only rescues and one unmitigated case. Partial mitigation must not be equated with elimination of harm. This taxonomy applies to the three core policies in seed 42; GCB baseline ties have separate equality handling.
Figure 9: Illustrative corrected seed-42 diagnostic. CASL versus Always-TL RMSE changes on the same 201 fixed-harmed wells, with equal well weight. Negative histogram values indicate reduced error; the median is −0.03799 . The scatter compares excess RMSE relative to No TL before and after blending: points below zero are full rescues and points below the diagonal receive partial mitigation. These are descriptive changes relative to Always TL, not evidence that CASL provides stronger protection than GCB.
Figure 10: Descriptive Cliff’s delta for the same previously available well descriptors and seed-specific outcome groups. Positive values denote larger descriptor values in the left group. Train–test shift uses realized test truth; source similarity uses training histories. Seed agreement describes outcome sensitivity to training randomness on the same cohort, not independent ecological replication or causal evidence. Magnitude/shift full-rescue-versus-persistent contrasts repeat direction in four seeds; recent change repeats in five. Within-fixed-harm Spearman associations, discussed separately in the text, are a different estimand. Seed 23 has no newly harmed wells and no right-panel contrast; remaining new-harm groups contain only one to nine wells. The seed-42 new-harm profile is illustrative rather than a general mechanism. These comparisons do not establish relative predictive importance or added protection from contextual features.
Policy
RMSE
MAE
No TL
4.488063
1.949304
Always TL
4.477892
1.906858
Global Convex Blend
—
—
CASL
4.476309
1.916308
Site-Hard
4.483016
1.909162
Appendix
Table 8: Corrected seed-42 retrospective errors on the 2,759 test observations with ∣wtda∣>1.5 . These anomaly-magnitude groups are not validated drought events. GCB stratum errors were not computed in the saved analysis and remain unreported. No event-detection metrics are implied.
Figure 11: Illustrative corrected seed-42 diagnostic. Exploratory distributions for full-rescue (88), persistently harmed (113), preserved-benefit (247) and newly harmed (nine) wells. Panels distinguish forecast-origin inputs, evaluation support, test-truth retrospective diagnostics and BRO registry snapshots. Origin anomaly magnitude, endpoint change, train–test shift and screen depth use symmetric logarithmic axes to retain all outliers. Boxes show median/interquartile range and points retain observations; missing diagnostics remain missing. The small new-harm group and multiple explored features preclude causal or confirmatory interpretation. These distributions do not establish relative predictive importance. Screen metadata are not established historical model inputs.
Figure 12: Illustrative corrected seed-42 diagnostic. Left: blending weight γ over 13,718 corrected test observations. Weights remain within [0,1] without collapse to either parent. Right: mean γ per well by outcome group. The γ≥0.5 convention describes dominant blend weight, not hard routing, event probability or a cause of group membership. Gate fitting used only 6,562 calibration rows. Absence of collapse does not establish superior harm protection.
Figure 13: Illustrative corrected seed-42 diagnostic. Inference-only sensitivity of the saved hgb_env gate under joint permutation of nine disjoint input groups covering all 47 features. Dots show mean equal-well-weight absolute change in γ over eight permutations; lines show the 5th–95th percentile randomization spread. No estimator is fitted and no test truth enters this calculation. Group size and correlated inputs affect sensitivity, so this is neither conditional importance nor proof of physical mechanism, improved predictive accuracy or added protection over GCB.
Reliable forecasting of multivariate time series under anomalous conditions is crucial in applications such as ATM cash logistics, where sudden demand shifts can disrupt operations. Modern deep forecasters achieve high accuracy on normal data but often fail when distribution shifts occur. We propose Weighted Contrastive Adaptation (WECA), a Weighted contrastive objective that aligns normal and anomaly-augmented representations, preserving anomaly-relevant information while maintaining consistency under benign variations. Evaluations on a nationwide ATM transaction dataset with domain-informed anomaly injection show that WECA improves SMAPE on anomaly-affected data by 6.1 percentage points compared to a normally trained baseline, with negligible degradation on normal data. These results demonstrate that WECA enhances forecasting reliability under anomalies without sacrificing performance during regular operations.
Joel Ekstrand, Tor Mattsson, Zahra Taghiyarrenani +3
Center for Applied Intelligent Systems Research, Halmstad University, Halmstad, Sweden · Mikael Linden Consulting AB, Stockholm, Sweden
Source selection often favors similar data, yet marginal proximity need not identify the most useful pretraining source. We motivate a conditional alternative through domain-adaptation theory and a regression example in which broader source variation improves estimation despite increasing Wasserstein distance. We examine five-day cryptocurrency volatility using ten Binance spot assets, six forecaster families, three seeds, and shared temporal cutoffs. Selecting the three farthest rather than nearest Gaussian-Wasserstein sources improves mean persistence-relative skill by 12.52 percentage points for MLP and 20.90 for Transformer, with positive post-hoc paired date-block intervals. Forward pilot selection improves mean skill over random transfer in all six families. Compact recurrent models achieve the lowest absolute mean errors, separating selection utility from forecasting quality. Retrospective five-candidate screening shows additional gains from farthest-first geometry before pilot ranking for MLP and Transformer. We distinguish across-target associations from within-target decisions and evaluate target heterogeneity, training seeds, and temporal-block sensitivity. Findings are exploratory for a fixed historical panel; the screening analysis measures no runtime saving.
Gordei Verbii, Georgii Killinkarov, Vladimir Yafarov
School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China · Stratwise, Singapore
Contextual specialization can improve forecasting accuracy, but a correction selected on one historical interval may become unreliable under temporal distribution shift. To address this issue, we propose RAEGL, a Risk-Aware Evidence-Gated Learning framework for selective contextual forecasting. RAEGL retains a validated global predictor by default and activates a contextual residual only when pre-deployment evidence supports its use. The framework separates candidate selection from gate calibration and jointly evaluates randomization significance, practically meaningful gain, and temporal stability. Experiments on real-world audits and controlled panels show how RAEGL can prevent harmful contextual deployment while making conservative opportunity costs explicit. In a reconstructed Our World in Data audit, exact fallback avoids RMSE degradations of 0.0960 and 0.0239 caused by two validation-selected corrections. In a sealed World Development Indicators evaluation, a region-based correction passes the randomization test but is withheld because its gain is only 0.000092, its country-clustered 95% confidence interval crosses zero, and only 0.02% of bootstrap replicates reach the practical threshold. In controlled panels, the stability- and support-aware extension activates in 97.2% of strong, stable-context runs while rejecting all high-drift settings. These results support RAEGL as an auditable, evidence-based mechanism for managing contextual deployment risk and as a conservative alternative to validation-driven contextual selection.