Executive Summary

Key numbers

All four figures are ranges taken across five repeated runs. The fact that they do not settle on a single value is itself part of what the paper reports. Hold the configuration fixed and change nothing but the random seed, and each of them still lands anywhere inside the ranges below.

Source: Selective Transfer Learning for Groundwater Level Anomaly Forecasting, arXiv 2610.37733.

The paper puts four methods side by side. All of them use the same network architecture, and the only thing that changes is how much of the outside learning each one lets in.

• • No transfer (No TL). Trained from scratch on the Dutch well records alone. This is the floor everything else is measured against
• • Always transfer (Always TL). Pretrained on hydroclimatic gridded data, then fine-tuned on the well records. This is what the industry usually means by transfer learning
• • Global Convex Blend (GCB). Mixes the predictions of the two models above, with the mixing weight fixed at one value for the whole run. Every well uses the same weight
• • Context-aware selective learning (CASL). Mixes the same two models, but sets the weight separately for each well and each point in time. This is what the paper proposes

Averaged over five repeated runs, the scoreboard comes out as follows. Lower is better for the two error measures, RMSE and MAE, and higher is better for R², which reports how much of the variation the model explains.

Source: arXiv 2610.37733, Table 2. Means and sample standard deviations over five seeds.

Read the table alone and the four are much the same model. The RMSE gap between the best and the worst is 0.016, and the spread attached to each value is as wide as that or wider. In a meeting where the average is what gets presented, this would have closed with one line: transfer learning was applied and the error improved slightly.

Go column by column and the leader changes. On MAE, which averages the size of the errors as they are, always transfer sits lowest at 0.902 and the selective method follows at 0.904. On RMSE, which penalizes large errors more heavily, and on R², the order flips and the selective method leads. The paper describes the best model depending on the loss function as a consistent pattern across this comparison. None of the gaps clear the spread, so no ranking can be nailed down, but the table already carries the point that picking a method off one column gives a different answer than picking it off another.

Split the result by well and the picture changes. Across five runs, always transfer left 18.38 to 51.86 percent of the wells worse off than the model with no pretraining. In the worst case, seed 42, that was 201 of the 457 wells, or 43.98 percent. The swing matters as much as the level: nothing changed between runs except the random seed, and the size of the damage still moved by more than a factor of two. Whether one well in five takes the loss or something close to half of them cannot be known in advance.

The authors do draw a boundary around what that swing represents. The five runs hold the data and the split fixed and vary only the randomness of training, so the 18.38 to 51.86 percent range describes training randomness. How much further the spread would open on a different period or a different population of wells is not a question this range answers.

The subjects are 457 groundwater monitoring wells in the Netherlands. What the model has to predict is the anomaly, meaning how far the groundwater level departs from its normal seasonal value. The input is nine consecutive monthly observations with no gaps, and the target is the value exactly three calendar months later. No interpolation was used to fill missing months. Patching a gap to force a usable window would leave the model predicting smoothness it created itself, and that door was shut at the start. What survived the rule is 13,718 test windows. Training runs on records up to July 2008, and testing is held out from June 2013 to June 2016.

Look between those two dates and the timeline is cut into four pieces. Training uses October 1996 through July 2008, validation runs from August 2008 to June 2011, calibration from July 2011 to May 2013, and testing from there through June 2016. The third piece, calibration, is set aside purely for fitting the selector that appears in section 4. Because the four pieces follow each other in time, the design closes off any path by which a model trained on an earlier stretch could see a later one.

The outside data, the source domain, is three hydroclimatic gridded products: TSMP-G2A, COSMO-REA6 and GLEAM. Those grids yield 122,880 rows across 512 cells, of which the researchers kept four. Selection worked by computing the cosine similarity between trajectories over the overlapping period and keeping only cells above 0.60, after which the pool was fixed once on seed 42 using the validation error of always transfer alone and held constant through all five runs.

One thing is worth stating here. Source and target are not the same kind of measurement. One side is model output averaging over a wide grid cell; the other is the water level at a single point where a well is sunk into the ground. They carry related signals without sharing a ruler. The authors write that this mismatch may contribute to how differently wells turn out, while noting that their descriptive comparisons do not establish that mechanism.

The first safeguard that comes to mind when bringing in somebody else's model is to pick data that resembles your own. A model trained on records that look like your site should go wrong less often, the thinking goes, and the paper applied a similarity threshold of its own when choosing source cells.

Compute the relationship between similarity and transfer outcome separately, though, and the expectation wobbles. Across all 512 grid cells the correlation ran from −0.080 to 0.054, and within the four cells that passed the threshold and were actually used it ran from −0.056 to 0.072. It is not only small; the sign turns over. There is no support for saying that closer resemblance helps, and none for saying it hurts. The authors themselves state that these comparisons do not establish relative predictive importance, causal mechanisms or prospective failure detection.

This is what negative transfer looks like in the field. When the representation learned on the source pulls against the patterns in the target, what was learned in advance becomes an obstacle rather than a head start. The trouble is that the pull barely registers in an average. Wells that gain and wells that lose cancel each other out, and the pooled error sits quietly where it was.

The route the researchers took is to ask the transfer question again at every well. Both parent models, no transfer and always transfer, are trained first and then frozen, and a separate gate is trained to decide how much of each prediction to mix. The gate reads 47 inputs: recent level state, calendar information, the sine and cosine terms that encode season, site history such as the mean and variability of that well over the training period, and summaries of similarity to the source. The gate is a histogram gradient boosting model, and it emits a mixing weight between 0 and 1 for every single observation. The calibration data used to fit it is 6,562 observations drawn from 407 wells.

Those 6,562 observations all come from the calibration stretch between July 2011 and May 2013. Within that stretch, 50 wells yielded no eligible window at all. So 407 of the 457 wells took part in fitting the gate, and the remaining 50 stayed in the test cohort without a single row of their own going in. The design promises a decision tailored to each site, and for more than a tenth of the sites there was nowhere to gather the basis for that decision.

The effect shows up most clearly where the damage was. Of the wells harmed by always transfer, 97.53 to 99.50 percent lost less under the selective method. Losing less and not losing are different things, though. The wells carried all the way past the no-pretraining baseline numbered 23.46 to 43.78 percent. Take seed 42: of the 201 harmed wells, 200 improved at least somewhat, and 88 of those cleared the baseline. Running the other way, 3.52 percent of the wells that had been gaining under always transfer were newly harmed by the selective method.

Sorting all 457 wells of seed 42 into five mutually exclusive outcomes sharpens the shape of those proportions: 88 full rescues, 112 wells where the loss merely shrank, one with not even that, 247 that were gaining under always transfer and kept the gain after selection, and nine newly harmed. Read only the harm and recovery figures in sequence and transfer looks like a wrecked experiment, when the largest block is the 247 that gained from the start and never gave the gain back. That is also why the paper does not tell anyone to stop transferring.

Here is the caveat that sets this paper apart from other work on transfer learning. Convexly blending two predictions shaves off extremes by construction. Add even a little of a safe prediction to one that has swung badly and the error drops. So a partial recovery may come out of the blending operation itself rather than from the gate having recognized anything about context. The authors put that sentence in the abstract.

The same seed shows what separates the selective method from the fixed blend. Under the selective method, 113 wells stayed harmed and nine were newly harmed, leaving 122 on the losing side; the Global Convex Blend left 124, with no new harm at all. Two wells apart in total, and the character of the harm differs. Cutting the weight finely can manufacture a loss that was not there, and tying it to one value leaves no room for that.

Neither method is declared the winner. In four of the five runs the fixed blend protected more wells, and in four the selective method had the lower pooled error. Accuracy and risk trade against each other here; neither one contains the other. The scope has lines drawn around it as well. Monthly data flattens the extremes of a drought that passes quickly, the rolling redeployment used in practice was not validated, and whether the model actually catches drought events was not evaluated. Because the selective method and the fixed blend are calibrated against different objectives, isolating what the context inputs contributed is not possible either.

Strip away the wells and the remaining structure is familiar. Bringing an externally trained model into a site that is short of data happens the same way in anomaly detection on factory equipment, in reading medical images, and in demand forecasting for stores whose customers differ branch by branch. The evidence anyone holds at that moment is usually one average: a benchmark score went up, or a pilot cut the error by a few percent.

What the Amsterdam result shows is how much that single line can cover. While the average over 457 wells hardly moved, at least one site in five inside it was losing ground. Whoever receives the average has no way to tell which side their own site is on. So before an externally trained model goes anywhere near your data, three things are worth settling.

• • Ask for the distribution next to the average. Not how many percent the error fell, but how many sites out of the total got worse, written into the same report. The paper's Table 2 does not tell the four methods apart; the difference only appears once the results are broken out at the level of individual wells
• • Do not treat a similarity score as a warranty. A number saying source and target resemble each other can inform the choice without guaranteeing the outcome. In this data even the sign of the correlation refused to hold still. We picked similar data, so we are safe is an expectation, not a finding
• • Keep a record of the weight you mixed at. The authors recommend reporting the mixing weight alongside the forecast. Knowing how heavily a given site leaned on the external model is what later makes it possible to ask whether the data or the weighting was at fault when a result goes wrong

The paper draws a line under that third item. The weight recorded next to a forecast lets a downstream user audit which parent contributed more, and that is all it does: it is not a probability of transfer safety and it is not a physical-causality model. Keeping a record and being assured of safety sit in different columns on this point as well.

When we talk about AI-Ready Data we mostly mean the work of putting data into a state a model can use. This study adds a question to that. What will you use to confirm that the model you brought in fits your data? The average scoreboard did not answer it, and neither did the similarity score. What remains is breaking the results out site by site, and keeping a record of what the breakout showed.

Thanks for reading this far. The figures and quotations here were checked against the abstract and body of arXiv paper 2610.37733. The paper is a preprint that has not been through review, and it was never tested the way a deployed model is, with periodic refits as time moves forward, so please read it with that in mind. If you have brought an externally trained model into your own sites and watched the score fall at only some of them, we would like to hear what tipped you off.

Pebblous Data Communication Team
November 1, 2026

Academic papers

• 1.Wu, Y., Alsahag, A. M. M., & Mohammadi Ziabari, S. S. (2026). "Selective Transfer Learning for Groundwater Level Anomaly Forecasting: Accuracy-Risk Trade-offs." arXiv:2610.37733.