PaperlayerView exact revision on arXiv

Source-linked AI summary

Training on the Future: A Delay-Aware Audit of Test-Time Adaptation for Time-Series Forecasting

Mohamed Readh Fentazi, Mazene Ameur, Adlen Ksentini

arXiv:2610.12232v1cs.LG

TL;DR

Forecast labels arrive only after the H-step target window matures, and real pipelines can delay their use further. This paper audits published adapters and closed-form correctors under explicit causal label delays. Across five benchmarks, the closed-form bar is cleared only on ETTm1, while apparent gains can change substantially under leaky evaluation.

  • Problem

    For H-step forecasting, a label does not exist until H steps after issuance, and real data pipelines can delay usable labels further, complicating trustworthy TTA evaluation.

    Source: §1 · p. 1

  • Method

    The harness releases labels at step s+d and enforces that delay inside four published methods' update loops, evaluating them alongside an RLS-Bank and an ELF-style corrector.

    Source: §1 · pp. 1–2

  • Results

    Across five benchmarks, the audited stack clears the closed-form corrector bar only on ETTm1: DynaTTA beats the bank at minimum causal delay at roughly 2,500 times its per-update cost, while ELF remains ahead.

    Source: §1 · p. 2, §7 · p. 10

  • Takeaways & Limitations

    The authors recommend causal delay reporting, naive and frozen baselines, closed-form correctors, adaptation cost, and backbone training recipes in TTA evaluations.

    Source: §7 · p. 9

    The audit covers five dataset families but only PatchTST; DynaTTA is absent on Traffic, and Weather, Electricity, and Traffic use single checkpoints and a reduced external-method set.

    Source: §7 · p. 9

Abstract

from arXiv · show

Test-time adaptation (TTA) methods for time-series forecasting update a deployed model, or a small adapter around it, from incoming ground truth. But the label of an $H$-step forecast exists only $H$ steps later, and real data pipelines add further delay. We build a leakage-free harness in which the label of forecast origin $s$ is released for updates only at step $s+d$ with $d \ge H$, and enforce this rule inside the released code of four recent TTA methods (TAFAS, COSA, PETSA and DynaTTA), run on their own backbones and checkpoints across five benchmarks (ETTm1, ETTh2, Weather, Electricity and Traffic). As references we add two closed-form correctors: a bank of recursive least squares (RLS) filters combined by a per-coordinate median, with no tunable hyperparameters and 56 microseconds per step on the 7-channel streams, and an ELF-style linear corrector. Under causal delayed labels the picture is asymmetric. On ETTm1 every audited method genuinely adapts, yet the RLS bank still beats three of the four at a fraction of their cost; only DynaTTA beats the bank, only at the minimum causal delay, and at roughly 2,500 times the per-update cost; the ELF-style corrector beats all four. On the other four datasets, the largest statistically significant improvement any published method achieves over its own frozen checkpoint is half a percent, on all four at least one published method is significantly worse than the frozen model at the minimum causal delay, and on drift-heavy ETTh2 longer label delays make every adapter that separates from the frozen model, ours included, significantly harmful. Leaky next-step updates inflate the apparent gains of simple adapters by up to 110%, and the backbone training recipe moves frozen online error by up to a factor of 25, more than any adaptation effect we measure. We release the harness, integration patches and all cached runs.

1 INTRODUCTION

The paper makes label delay an explicit part of test-time adaptation evaluation and audits published methods alongside closed-form correctors. Its results are bounded: the audited stack clears the closed-form bar only on ETTm1, while evaluation practice can substantially change apparent gains.

1 INTRODUCTION

  • The harness releases each forecast-origin label only at step s+d, with d=H as minimum causal delay and d>H modeling pipeline latency.The d=1 legacy setting is retained only as a leaky reference.Source: §1 · p. 1
  • TAFAS, COSA, PETSA, and DynaTTA are evaluated from their own repositories and checkpoints across five benchmarks, with delay enforced inside their update loops.Their partial-ground-truth mechanisms use causally shortened prefixes.Source: §1 · p. 2
  • The RLS-Bank combines filters across forgetting factors by a per-coordinate median, has no tunable deployment hyperparameters, and costs 56 microseconds per step on a 7-channel stream.The ELF-style corrector is also evaluated under released-label updates.Source: §1 · p. 2
  • On ETTm1, the RLS-Bank beats three of four audited methods and the ELF-style corrector beats all four; outside ETTm1, significant gains by published methods top out at 0.5%.On each of the other four datasets, at least one published method is significantly worse than its frozen checkpoint at the minimum causal delay.Source: §1 · p. 2
  • The study reports that gradient adapters are fragile in ways closed-form updates are not, with up to eleven of twelve gradient configurations diverging per cell.Source: §1 · p. 3
  • The paper does not claim gradient adaptation cannot help: all four audited methods help on ETTm1, and DynaTTA beats the bank there.Its claim is that closed-form correctors clear the audited stack's bar exactly once across five datasets.Source: §1 · p. 3

2 RELATED WORK

Prior work spans forecasting-specific test-time adaptation, lightweight correctors, classical adaptive filtering, and causal online protocols. This paper connects these lines by varying label delay while auditing published methods and closed-form baselines under causal evaluation.

Test-time adaptation for forecasting

  • Forecasting-specific TTA includes gated calibration, parameter-efficient adapters, single gated output adapters, shift-aware adaptation, and frequency-domain calibration, but prior work does not jointly vary delay, include closed-form competitors, and report naive baselines causally.The methods include TAFAS, PETSA, COSA, DynaTTA, and FAC.Source: §2 · p. 3

Frozen models with lightweight correctors

  • ELF fits an OLS-equivalent linear forecaster online beside a frozen foundation model and combines them with online weighting; concurrent ORCA uses replay-buffer training with a maturity rule.Neither line varies delay or audits the TTA stack.Source: §2 · p. 3

Classical adaptive filtering

  • Classical adaptive filtering runs RLS filters with different forgetting factors in parallel, but learned mixing weights rely on feedback that the paper's ablations find breaks under delayed labels.Delayed RLS updates have a history in applied work, including k-step stale-regressor updates.Source: §2 · p. 3

Causal protocols and delayed feedback in online forecasting

  • Causal online-forecasting protocols remove leakage, while other work studies delayed feedback; this paper extends that line to test-time adaptation by varying d and enforcing it in published code.It also reports naive baselines in every table.Source: §2 · p. 3

3 THE CAUSAL DELAYED EVALUATION PROTOCOL

The protocol treats label availability as an experimental variable, distinguishing fully matured labels from additional pipeline lag and leaky updates. It evaluates frozen-backbone adapters on a long online stream with aligned predictions and paired uncertainty estimates.

3 THE CAUSAL DELAYED EVALUATION PROTOCOL

  • The five benchmarks comprise ETTh2 and ETTm1 with 7 channels, Weather with 21, Electricity with 321, and Traffic with 862; each stream uses a 25%/5%/70% pretraining/validation/online split.Standardization uses pretraining statistics only.Source: §3 · p. 3
  • A target window from origin s matures at s+H and is released at s+d: d=H is the minimum causal delay, d>H adds lag, and d=1 is leaky.Additional lag also shortens the usable partial-ground-truth prefix.Source: Figure 2 · p. 4, §3 · p. 4
  • Each frozen backbone maps a 96-step lookback to a 24-step forecast at every stream step, and adapters may modify forecasts but never update the backbone.Internal comparisons replay the same frozen stream; the external audit uses the authors' checkpoints and recipes.Source: §3 · p. 4
  • All four audited methods are gradient-based and retain their distinct adaptation designs, including partial-ground-truth mechanisms whose prefixes are causally shortened as lag grows.TAFAS uses gated calibration and periodicity-aware scheduling; PETSA uses low-rank adapters; COSA uses a gated output module; DynaTTA adapts in input space based on measured shift.Source: §3 · p. 4
  • Performance is standardized MSE over each full online stream, with significance assessed by paired per-step loss differentials and a circular moving-block bootstrap.A difference is significant when its 95% interval excludes zero; internal PatchTST cells use three seeds on ETTh2, ETTm1, and Weather, and one on Electricity and Traffic.Source: §3 · p. 4
  • The audit applies label-lag patches in the methods' update loops, exports per-origin predictions, and aligns forecasts before scoring identical tensors.DynaTTA is omitted on 862-channel Traffic because its input-space update is prohibitively slow.Source: §3 · pp. 4–5

4 THE RLS-BANK

The RLS-Bank corrects each frozen forecast using channel- and horizon-specific filters, then combines experts without relying on delayed feedback. Its feature centering and covariance clip address instability, while the median handles regime variation and failing experts.

4 THE RLS-BANK

  • For each channel and horizon step, the correction uses the backbone forecast, an intercept, and two deviation features centered by the trailing daily mean; weights start at passthrough.At initialization, w=[1,0,0,0], so the corrected forecast matches the frozen forecast.Source: §4 · p. 5
  • Centering reduces covariance wind-up risk: raw level-tracking features can be strongly correlated, allowing the RLS covariance matrix to grow in unexcited directions.The features include the last observation and seasonal-naive value, each centered by the recent-day mean.Source: §4 · p. 5
  • Each released matured sample triggers an exponentially weighted RLS update, followed by covariance symmetrization and a trace clip that bounds gain and prevents wind-up.The clip is not a learning rate; it costs at most 2.8% higher MSE on narrow streams and separates working from exploding behavior on the 321- and 862-channel streams.Source: §4 · p. 5
  • Because the best forgetting factor varies by dataset, backbone, and delay, the bank takes the per-coordinate median of six RLS experts and a frozen passthrough expert.Feedback-driven selection can lag a failing expert by up to d steps; the median needs no feedback and absorbs extreme experts on wide streams.Source: §4 · p. 5

5 EXPERIMENTS

The experiments compare closed-form and gradient adapters on an internal grid and audit published methods on their own checkpoints across five benchmarks. Under causal delays, published-method gains are concentrated on ETTm1; elsewhere, improvements are small or methods can harm frozen-model performance, while cost and training-recipe choices materially affect the comparison.

5 EXPERIMENTS

  • The internal grid spans five datasets, two backbones, four delays and up to 34 methods, while the external audit runs four published methods on their own checkpoints.The internal grid contains 2,144 cached runs; external audits cover all five datasets, with three checkpoint seeds on ETTh2 and ETTm1.Source: §5 · pp. 5–6

5.1 CONTROLLED GRID: CLOSED-FORM VERSUS ORACLE-TUNED GRADIENT ADAPTERS

  • One closed-form corrector is lowest in all ten internal cells; RLS-Bank matches or beats oracle gradients in six, with deep ETTh2/PatchTST the sole significant loss.The internal cells compare update rules on deliberately imperfect backbones; conclusions about published methods come from their own checkpoints.Source: §5.1 · p. 6, Table 1 · p. 6
  • On ETTh2, causal bias-EMA MSE exceeds its leaky-update MSE by 83% with DLinear and 110% with PatchTST.At d=1, several adapters appear to beat persistence, but none does at d=H.Source: §5.1 · p. 7

5.2 PUBLISHED METHODS ON THEIR OWN CHECKPOINTS

  • On ETTm1, all four published methods adapt, but RLS-Bank beats three; DynaTTA alone beats the bank at d=H, while ELF-mix beats all four.DynaTTA's win over the bank is 7% and costs roughly 2,500 times more per update.Source: §5.2 · p. 7, Table 2 · p. 7
  • Outside ETTm1, the largest significant published-method gain is TAFAS's 0.5% on Traffic; all four methods worsen Electricity, and ETTh2 adapters that separate from frozen become harmful as delay grows.The reported Electricity comparison is at d=H; ETTh2 harm appears from d=2H for the bank and at d=4H for several published methods.Source: §5.2 · p. 8

5.3 COST

  • On ETT streams, RLS-Bank updates cost 0.056 ms per step, two to four orders of magnitude less per update than the compared gradient methods.Its cost averages 2.9 ms per step across all ten internal cells, including Traffic.Source: §5.3 · p. 8

5.4 DIAGNOSTICS: LABEL STALENESS, PLASTICITY AND THE TRAINING RECIPE

  • As labels grow staler, the winning gradient learning rate shrinks and the RLS memory grows; the bank stays within 7% of the per-cell oracle in nine of ten cells.Weather/DLinear at d=H is the exception, where the bank is 24% above the best single forgetting factor.Source: §5.4 · p. 8
  • PatchTST's training recipe changes frozen online error by a factor of 5 on ETTh2 and 25 on Electricity, exceeding any measured adaptation effect.On ETTm1, the direction reverses: the paper's recipe scores 1.14 versus 1.52 for the authors' recipe.Source: §5.4 · pp. 8–9

6 WHEN CLOSED-FORM ADAPTATION FAILS

Closed-form adaptation can fail when delayed feedback makes most experts mildly harmful, and its advantage does not grow with label delay. The section also identifies settings where gradient or ELF-style correction performs better.

ETTh2: the median follows a poisoned majority

  • On ETTh2, the RLS bank is significantly harmful at d = 2H (+0.16 to +0.26) and d = 4H (+0.33 to +0.44), as are COSA, DynaTTA and PETSA at extended delay.The median blocks a minority of exploding experts, not a majority of mildly harmful ones; ELF-mix’s frozen-model anchor instead converges to the frozen forecast.Source: §6 · p. 9

Weather: the median follows a poisoned majority

  • On Weather/DLinear at d = H, the RLS bank is 24% above the best single forgetting factor because most experts are mildly poisoned, then recovers at longer delays.The median follows the majority in this regime.Source: §6 · p. 9

A hindsight-tuned gradient method can win on heavy tails

  • On ETTh2/PatchTST at d = H, plain SGD beats RLS significantly by 0.38, showing that hindsight-selected gradient learning rates can win under very heavy-tailed residuals.Per-direction gain control alone is not enough in this setting.Source: §6 · p. 9

The closest published relative beats us on two of the three streams where it runs

  • The ELF-style corrector is significantly better than the bank on ETTm1 by about 0.10 to 0.16 and on ETTh2 by 0.13 to 0.61 across seeds and delays, and is statistically tied on Weather at minimum delay.The bank retains cost and wide-stream applicability advantages, where ELF-style correction is impractical.Source: §6 · p. 9

The closed-form advantage is robust to delay, not increasing in it

  • The closed-form advantage is robust to delay, not increasing in it, contrary to the paper’s initial hypothesis.On ETTh2, the bank becomes harmful at extended delay.Source: §6 · p. 9

7 LIMITATIONS, RECOMMENDATIONS AND CONCLUSION

The authors delimit the audit’s scope, recommend reporting causal delay, baselines, cost, and training recipe, and conclude that the audited stack clears the closed-form bar only once. The appendices document integration, dependency and alignment choices, plus released reproducibility artifacts.

7 LIMITATIONS, RECOMMENDATIONS AND CONCLUSION

  • The audit covers five dataset families but one deep architecture, PatchTST; the ELF-style corrector is limited to C ≤21, leaving wide streams without that reference.Other scope constraints include reduced seeds or method sets on some streams and a deliberately nonstandard 25/5/70 split.Source: §7 · p. 9
  • The authors recommend causal d = H and at least one d > H, reporting persistence, seasonal-naive and frozen baselines, closed-form correctors, adaptation cost, and backbone training recipe.On the five benchmarks, the audited stack clears the closed-form bar only with DynaTTA on ETTm1 at minimum causal delay, at roughly 2,500 times the per-update cost; ELF remains ahead.Source: §7 · pp. 9–10
  • The paper reports use of generative AI for implementation and selected writing and figure tasks, with authors reviewing assisted work and taking responsibility for the final content.It reports that research direction, experiment design, execution and result verification were carried out by the authors.Source: AI USE STATEMENT · p. 10
  • The audit reuses public benchmark datasets, released code and checkpoints, and introduces no new datasets, human subjects or deployed systems.The authors report negative findings about existing methods and their own method under the same standards.Source: ETHICS STATEMENT · p. 10
  • The release includes cached runs, configuration hashes, a run manifest, harness notebook, integration kit, stream specifications and results, with reproducibility across three machines to 10−4.Appendices document deviations, corrector pseudocode and bootstrap procedures.Source: REPRODUCIBILITY STATEMENT · p. 10

A INTEGRATION AND DEVIATIONS LOG

  • The four audited methods run from the COSA reference repository on their own PatchTST checkpoints and training recipe, using the paper’s split and standardization; deviations are logged.Source: A INTEGRATION AND DEVIATIONS LOG · p. 12

A.1 DEPENDENCY INCOMPATIBILITY

  • The pinned NumPy and PyTorch versions were incompatible with the current Python/CUDA stack, so the authors installed the imported dependencies without replacing core libraries.All runner modules imported cleanly on that stack.Source: §A.1 · p. 12

A.2 PATCHES

  • The patches shift maturity triggers and shorten observed partial-ground-truth prefixes by label lag, export per-origin predictions, and fix compatibility and in-place autograd errors.At zero label lag, patched behavior is bit-identical to the authors’ code; the in-place fix enabled DynaTTA to complete adaptation events.Source: §A.2 · pp. 12–13

A.5 DATA LAYOUT

The appendix describes how benchmark data are laid out, clarifies the repository license, and relates delayed labels to nowcasting’s ragged-edge structure. It also cross-checks Figure 4’s counters using a second method’s run report.

A.5 DATA LAYOUT

  • The runner detects flat data files and creates the repository’s expected nested dataset layout; ETT data are fetched from ETDataset and the other benchmark files from the standard dataset release.Source: §A.5 · p. 13

A.6 LICENSE AND ATTRIBUTION

  • The COSA repository is CC BY-NC-SA 4.0, so the release distributes patches and integration scripts rather than modified repository copies.Source: §A.6 · p. 13

A.7 NOWCASTING CONNECTION

  • The paper’s label-lag setting shares the ragged-edge structure of econometric nowcasting, but its test-time adaptation setting and closed-form corrector are described as orthogonal to that literature.The paper makes no numerical comparison.Source: §A.7 · p. 13
  • DynaTTA’s run-report adaptation-rate histories—487 for ETTh2 and 1,951 for ETTm1—match COSA’s partial-ground-truth counters, independently confirming the Figure 4 values.Source: §A.7 · p. 14

B FULL RESULT TABLES

The appendix supplies full internal delay grids, external-checkpoint results, and adaptation-cost reporting, alongside the run-name conventions needed to read the tables. Captions specify dataset, backbone, delay, seed, and divergence conventions, but the supplied passages do not expose most table cells.

B FULL RESULT TABLES

  • The run-name key defines the frozen, persistence and seasonal-naive rows, auxiliary correctors, gradient adapters, and ∞/– markers for divergence or methods not run.It specifies ELF-style linear and mixed correctors as limited to streams with C ≤21.Source: B FULL RESULT TABLES · p. 15
  • Table 3 compares selected cells’ standardized MSE at leaky d=1 and causal d=H and reports relative increase; Tables 4–8 provide full internal grids.The d=1 columns are the leaky reference.Source: Table 3 · p. 15, B FULL RESULT TABLES · p. 15
  • Tables 4–6 give delay-specific standardized online MSE grids for ETTh2, ETTm1 and Weather, with DLinear on the left and PatchTST on the right; d=1 is leaky.Table 4 reports a three-seed PatchTST mean; the captions mark divergence and unrun cells.Source: Table 4 · p. 15, Table 5 · p. 16, Table 6 · p. 16
  • Tables 7–8 give Electricity and Traffic standardized online MSE grids across delays, with DLinear and PatchTST results; d=1 is the leaky reference.The captions identify one-seed PatchTST results and mark diverged or unrun cells.Source: Table 7 · p. 17, Table 8 · p. 18
  • Table 9 reports external evaluation on the published methods’ own PatchTST checkpoints for all five datasets, including auxiliary adapters replayed on frozen exports.ETT seed counts vary by method and delay; Weather, Electricity and Traffic use one checkpoint.Source: Table 9 · p. 18
  • Table 10 reports per-update wall time and adapted-parameter counts; update frequencies differ, so per-update ratios and amortized per-step costs are not interchangeable.The bank’s listed 4,032 closed-form weights apply to the 7-channel ETT streams.Source: Table 10 · p. 20

C ADDITIONAL FIGURES

The additional figures visualize delay effects, adapter sensitivity, label staleness, forecast-horizon error, cumulative error, and dataset drift across internal cells.

C ADDITIONAL FIGURES

  • Figure 3 compares standardized MSE across four label-release delays for frozen, persistence, RLS, RLS-Bank, and best-gradient-adapter references in ten internal cells.The ETTh2/PatchTST panel and both Electricity panels use logarithmic y-axes.Source: Figure 3 · p. 21
  • Figure 4 plots COSA and TAFAS online MSE against delay and reports COSA partial-ground-truth update counts, which fall to zero at positive lag under repository defaults.At zero lag, the counts are 487 for ETTh2 and 1,951 for ETTm1.Source: Figure 4 · p. 22
  • Figure 5 sweeps gradient-adapter learning rates and RLS forgetting factors at d=H, annotating feature-space divergences and reporting that all plain-SGD settings diverge on 321-channel Electricity.No RLS setting diverges on the 7- and 21-channel streams; on the two widest streams, the median bank absorbs individual large finite errors.Source: Figure 5 · p. 22
  • Figures 6–9 show oracle gradient learning rate and RLS memory versus delay, per-horizon MSE, cumulative-error curves, and rolling weekly means and standard deviations across dataset splits.Figure 6 notes that its decade-spaced learning-rate grid cannot display sub-decade movement.Source: Figure 6 · p. 23, Figure 7 · p. 23, Figure 8 · p. 24

D RLS-BANK DETAILS

This appendix specifies the RLS-Bank update and reports ablations on aggregation, input features, numerical clipping, and an ELF-style linear corrector.

D RLS-BANK DETAILS

  • At each released-label update, each channel-and-horizon filter performs an RLS update, then the bank emits the per-coordinate median forecast across its filters and passthrough expert.The algorithm initializes each filter with a passthrough weight and covariance matrix, and uses the released label indexed by j=i-d.Source: D RLS-BANK DETAILS · p. 26
  • Table 11 compares median aggregation with hindsight-best single-factor, follow-the-leader, and discounted follow-the-leader selection on five DLinear streams at d=H.The text identifies Electricity as the stream where selection wins because one fast forgetting factor is best throughout.Source: D RLS-BANK DETAILS · p. 26, Table 11 · p. 26
  • On ETTm1/PatchTST at minimum causal delay, rls feat scores 0.651 versus 1.138 for the frozen model, while rls calib scores 1.127.The calibration-only variant uses per-coordinate scale and shift; the two deviation features distinguish rls feat.Source: D RLS-BANK DETAILS · p. 26
  • On ETTh2/DLinear, every NLMS step-size setting yields 8.7–103 standardized MSE at causal and extended delays, versus 1.29 for the frozen model.The passage attributes the gap to NLMS lacking RLS’s covariance-based per-direction gain control.Source: D RLS-BANK DETAILS · p. 26
  • At d=H, unclipped single-filter error reaches 5.3 × 10^7 on Electricity/DLinear versus 5.96 clipped, while the unclipped bank survives everywhere by discarding exploding experts through its median.On 7- and 21-channel streams, clipping changes single-filter MSE by at most 0.4% on most cells, with a 2.8% improvement when unclipped on ETTh2/PatchTST.Source: D RLS-BANK DETAILS · p. 26
  • The ELF-style corrector fits raw-lookback-window recursive least squares and combines its forecast with the frozen model using an online inverse-error weighter.Both fitting and weighting update only on released labels; the method is restricted to streams with C ≤21 because its per-step cost is O(CHL2).Source: D RLS-BANK DETAILS · p. 27

E BOOTSTRAP DETAILS

The appendix describes paired moving-block bootstrap intervals, their interpretation and caveat, and the comparison tables for internal and external model evaluations.

E BOOTSTRAP DETAILS

  • Intervals resample paired per-step loss differences in circular blocks twice the seasonal period, using 500 resamples and reporting the 2.5th and 97.5th percentiles of the resampled mean.Negative mean differences favor the first method; an asterisk denotes a 95% confidence interval excluding zero.Source: E BOOTSTRAP DETAILS · p. 27, Table 12 · p. 27
  • The intervals may be anti-conservative because correlations at lags beyond the block length are not captured.For internal PatchTST cells with multiple seeds, paired differentials are pooled; the widest-stream internal comparisons cover only essential pairs at d=H and d=4H.Source: E BOOTSTRAP DETAILS · p. 27
  • Tables 12–15 cover internal-cell comparisons, external comparisons on authors’ ETTm1 and ETTh2 checkpoints, and external comparisons on Weather, Electricity, and Traffic checkpoints.External results are computed per checkpoint seed; Table 15 uses one checkpoint per dataset.Source: Table 12 · p. 27, Table 13 · p. 29, Table 14 · p. 30
  • For ETTm1/PatchTST at horizon 24, the mean difference for rls bank med minus frozen is -0.5012, with a 95% interval from -0.5765 to -0.4290.Tables 12–15 list cached intervals underlying the paper’s significance claims.Source: E BOOTSTRAP DETAILS · pp. 27–28

F RELEASED-SCRIPT CONFIGURATION RE-PASS (ETT STREAMS)

The appendix reruns TAFAS and COSA with released-script settings on seed-0 ETT streams and finds that configuration effects differ across ETTm1 and ETTh2.

F RELEASED-SCRIPT CONFIGURATION RE-PASS (ETT STREAMS)

  • The script settings change TAFAS’s learning rate to 0.001 and adjust COSA’s learning rate, steps, PAAS batch, context buffer, and PAAS flag from repository defaults.COSA’s context buffer grows from 5,047 to 5,887 parameters; larger batches preserve causal prefixes and produce 79 and 60 partial updates on ETTh2 at lags 24 and 72.Source: F RELEASED-SCRIPT CONFIGURATION… · p. 34, Table 16 · p. 34
  • Under the script configuration, COSA remains delay-degraded on ETTm1, despite the larger batches preserving causal-prefix partial-ground-truth updates at positive lag on ETTh2.The reruns use the same ETT checkpoints, streams, and origin-aligned scoring as the main text, with seed 0.Source: F RELEASED-SCRIPT CONFIGURATION… · p. 34
  • COSA’s script configuration is significantly worse than default by 0.15–0.32 on ETTm1 but significantly better by 0.76 on ETTh2 at d=4H; the bank uses one configuration everywhere.The script configuration is significantly worse on ETTm1 at every delay and equivalent to frozen on ETTh2 (|∆| ≤0.0005, n.s.).Source: F RELEASED-SCRIPT CONFIGURATION… · p. 34, Table 16 · p. 34

Related papers

Loading 2610.12232v1…