Source-linked AI summary
Beyond Wasserstein Proximity: Finding Better Transfer Sources
Gordei Verbii, Georgii Killinkarov, Vladimir Yafarov
TL;DR
Source selection often favors similar data, but marginal proximity may not identify the most useful pretraining source. The paper tests learner-specific selection rules for five-day cryptocurrency volatility; farthest-source selection beats nearest selection for MLP and Transformer, while forward pilot selection improves mean skill over random transfer in all six families.
Problem
For a fixed target, learner, and adaptation budget, it is unclear when a more distant source can be more useful than the closest source.
Source: §1 · p. 1Method
The study compares source-selection rules across ten Binance spot assets, six model families, and three seeds under shared calendar boundaries.
Source: §1 · p. 1Results
Farthest-over-nearest effects are 12.52 points for MLP and 20.90 for Transformer, while pilot selection improves mean skill over random transfer in all six families.
Source: §5.1 · p. 5, §7 · p. 7Takeaways & Limitations
Compact recurrent models achieve the lowest mean errors, and matched-budget screening suggests an incremental role for farthest-first geometry before predictive validation.
Source: §6 · p. 7, §7 · p. 7Ten assets and one test period limit generalization, and post-hoc pointwise intervals are unadjusted for multiplicity.
Source: §6 · p. 7
Abstract
from arXiv · showhide
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.
1 Introduction
Marginal similarity is a plausible but incomplete guide to source selection: for a fixed target and adaptation budget, more distant sources may help depending on the learner. The paper tests this conditional idea in cryptocurrency forecasting and separates selection utility from forecasting quality.
1 Introduction
- Similarity can tighten domain-adaptation bounds when other terms are controlled, but distance alone does not order sources with different predictive relationships or estimation quality.Matching feature marginals can also leave conditional differences unresolved.Source: §1 · p. 1
- A source with broader informative variation can estimate a shared response more accurately while remaining farther away in marginal geometry.The paper presents this as a mechanism to investigate, not one established by the cryptocurrency experiment; it does not imply a universal preference for dissimilarity.Source: §1 · p. 1
- For a fixed target and finite adaptation budget, more distant Gaussian-Wasserstein sources may yield lower forecast error than nearer ones, depending on the learner.The paper motivates this conditional hypothesis with domain-adaptation theory and a regression example in which broader informative source variation improves estimation despite greater marginal distance.Source: §1 · p. 1
- The paper evaluates ten Binance spot assets, six model families, and three seeds under shared calendar boundaries.Its contributions include comparing learner-specific source-selection rules, separating selection gains from absolute forecasting quality with controls, and testing geometric screening before predictive validation.Source: §1 · p. 1
2 Related work
Prior work motivates transferability measures based on discrepancy, predictive structure, and shared temporal components, but does not establish that a farthest-source rule is generally preferable. The paper positions its learner-specific comparisons against this background.
2.1 Domain adaptation and transferability
- Domain-adaptation bounds combine source risk, distribution discrepancy, and shared-hypothesis error, while task-aware transferability measures use target-label or feature–label evidence.Prior distant-task transfer work does not establish a farthest-source selection rule; this paper’s forward pilot selector directly measures adapted forecasting error on later target outcomes and serves as a validation comparator.Source: §2.1 · pp. 1–2
2.2 Dataset geometry and predictive similarity
- Gaussian Wasserstein distance compares means and covariances, but marginal distances can omit predictive structure and alternatives incorporate feature–label geometry or bidirectional generalization.The paper uses one-direction forward validation rather than the full Cross-Learning Score, and does not test superiority over these alternatives.Source: §2.2 · p. 2
2.3 Learning across time series
- Global and domain-adapted forecasters share temporal components in different ways, whereas this study uses separate single-source initializations followed by fixed target adaptation.Prior linear-model results motivate linear and compact recurrent controls but do not predict their ranking for five-day volatility.Source: §2.3 · p. 2
2.4 Volatility forecasting and evaluation
- The target is a logarithmic five-day RMS of daily returns, not the intraday realized-volatility quantity studied in cited econometric work.Proxy noise and loss choice can affect rankings, so findings concern the specified log proxy rather than an invariant ranking for latent variance or trading objectives.Source: §2.4 · p. 2
3 Why a more distant source can help
A regression example shows how broader source variation can improve estimation despite greater marginal distance, but only under a particular condition. The section also distinguishes pooled cross-target associations from within-target source-choice evidence.
3 Why a more distant source can help
- For a ≥ 1, greater source distance accompanies smaller expected excess target risk because broader variation estimates the shared slope more accurately and finite adaptation retains the advantage.For a < 1, moving farther instead reduces information and increases risk; the illustration assumes a shared linear response, independent samples, and population target updates, not cryptocurrency markets or neural fine-tuning.Source: §3 · p. 2
3.1 The relevant conditioning direction
- Across-target correlations for a fixed source answer a different question from varying sources for one target, and pooled associations can include between-target difficulty.The paper therefore uses within-target policy losses as operational evidence for source choice.Source: §3.1 · p. 2
4 Experimental design
The design fixes a historical multi-asset forecasting panel and compares geometric and pilot-based source selection under shared temporal evaluation. It also separates source-selection comparisons from retrospective screening and diagnostics.
4.1 Data, target, and information availability
- Daily Binance spot bars for ten assets support forecasts of standardized log-volatility from 60-day feature histories, with shared train, pilot, and test boundaries and five-day-spaced pilot and test origins.Source: §4.1 · pp. 2–3
4.2 Geometry and training
- Gaussian-Wasserstein geometry uses shrunk Gaussians over flattened 30-day feature windows, while pretraining uses each source’s eligible training history; six architectures receive fixed-epoch training and adaptation.Source: §4.2 · p. 3
4.3 Source choice, metrics, and uncertainty
- Three sources are selected by nearest distance, farthest distance, or pilot loss; random transfer is the all-source mean, effects use persistence-relative skill, and uncertainty uses paired date-block intervals.The figures also distinguish correlations across partners for each fixed asset from averages across assets; these descriptive distributions are not confidence intervals.Source: §4.3 · pp. 3–4, Figure 2 · p. 4
4.4 Retrospective screening and diagnostics
- Retrospective screening compares nearest, farthest, or balanced shortlists at q = 3, 5, 7, or 9 before choosing the pilot-best three, against matched random shortlists.The q = 5 analysis is post-hoc, uses already-fitted candidates, and measures no runtime saving; diagnostics also examine pilot-to-test loss relationships, regret, seeds, and block lengths.Source: §4.4 · p. 4
5 Results
Farthest Gaussian-Wasserstein sources outperform nearest sources for MLP and Transformer, while forward pilot selection improves mean skill over random transfer in all six families. Compact recurrent models have the lowest macro MSE, and farthest-first screening adds gains for MLP and Transformer; ranking and screening findings remain exploratory for this fixed panel.
5.1 Operational source-selection effects
- Farthest-𝑊 selection beats nearest by 12.52 points for MLP and 20.90 for Transformer; recurrent-family effects are smaller and mixed.Source: §5.1 · pp. 4–5
- Pilot selection improves mean skill over random transfer in all six families, with gains ranging from 3.05 points for LSTM to 11.73 for Transformer.Source: §5.1 · p. 5
5.2 Across-target association and fixed-target choice
- Across-target mean distance–loss correlations are negative for every family, but within-target means are positive for three recurrent families, consistent with target heterogeneity.Source: §5.2 · p. 5
5.3 Absolute forecast quality and controls
- GRU-32 has the lowest pilot-selected macro MSE at 0.7653, while MLP and Transformer remain worse than HAR in mean error despite selection gains.Source: §5.3 · p. 5, Figure 3 · p. 5
5.4 Screening candidates before pilot selection
- At 𝑞= 5, farthest screening followed by pilot selection beats matched random-five screening by 4.24 points for MLP and 6.05 for Transformer; the retrospective analysis measures no runtime saving.Source: §5.4 · p. 6, Figure 4 · p. 6
5.5 Ranking stability and sensitivity
- Pilot–test ranking stability varies, and seed and block-length checks support stability only under tested choices, not replication on new assets or regimes.Source: §5.5 · p. 6
6 Discussion and limitations
The discussion separates source-selection utility from forecasting quality and treats the findings as exploratory. The fixed panel and evaluation design limit generalization, motivating broader prospective tests.
6 Discussion and limitations
- Compact recurrent models attain the lowest mean errors, while farthest-first screening is promising for MLP and Transformer.Possible contributors to the market-experiment results include source-history length, response compatibility, optimization, and regularization.Source: §6 · p. 7
- Ten assets and one test period limit generalization; fixed epochs leave unequal data and optimizer budgets, and recent-history geometry is not disentangled from full-history pretraining.The study uses 45 shared temporal origins.Source: §6 · p. 7
- Intervals omit retraining and reselection uncertainty, and pointwise intervals are unadjusted for multiplicity.Log-proxy MSE does not establish variance calibration, trading profitability, or intraday-volatility accuracy.Source: §6 · p. 7
- Prospective studies should rotate regimes, hold out asset cohorts, match training budgets, and vary adaptation length.Further comparisons could add task-aware distances, horizons, tuned GARCH benchmarks, measured screening costs, and multi-source learning.Source: §6 · p. 7
7 Conclusion
The conclusion treats marginal similarity as only one component of transfer utility and reports learner-specific selection results in a fixed panel. It also motivates prospective tests across assets, regimes, and computational budgets.
7 Conclusion
- Farthest-source selection beats nearest selection for MLP and Transformer, while forward pilot selection improves mean skill over random transfer in all six families.Compact recurrent models attain the lowest mean errors; matched-budget screening suggests an incremental role for farthest-first geometry before predictive validation.Source: §7 · p. 7
- A finite-adaptation example shows how a more distant but informative source can help, motivating a conditional hypothesis for cryptocurrency forecasting.The conclusion presents these findings as support for learner-specific source selection, not a universal preference for dissimilarity.Source: §7 · p. 7
- The findings motivate prospective tests across assets, regimes, and computational budgets.The reported conclusions are based on a fixed panel.Source: §7 · p. 7