Skip to main content

Beating Saturated SOTA the Lazy Way: Leakage-Free Online Ensembles for Long-Term Forecasting

·408 words·2 mins

TL;DR. The long-term time-series forecasting (LTSF) benchmark is saturated: PatchTST, iTransformer, TimesNet, TimeMixer, and even a linear model sit within a couple percent of each other. Instead of inventing model #47, we combine them with a leakage-free online ensemble and beat the best single model on 20/20 ETT+Exchange cells (statistically significant on 15/20) — with a parameter-free combiner that needs no tuning. The same method beat published N-BEATSx on all 5 electricity-price markets (see the previous post). This is a companion log to that one.


1. What is LTSF, for newcomers?
#

Long-term time-series forecasting means: given the last \(L\) time steps, predict the next \(H\) steps — where \(H\) is long (96 to 720 steps). The field’s standard playground is a handful of datasets:

  • ETT (Electricity Transformer Temperature): 7 sensors, 4 variants (ETTh1/h2 hourly, ETTm1/m2 every 15 min).
  • Exchange: 8 countries’ daily exchange rates.
  • Electricity (ECL, 321 channels) and Traffic (862 channels): the big, high-dimensional ones.

The standard rules (“Look-Back-96 protocol”): input length \(L=96\), horizons \(H\in\{96,192,336,720\}\), score by Mean Squared Error (MSE). Newcomer vocabulary: a channel is one variable/series; a rolling window is one (input → forecast) example slid along time; the test set is a contiguous later stretch evaluated as many overlapping windows.

The leaderboard is a parade of architectures — Autoformer → FEDformer → DLinear → PatchTST → TimesNet → iTransformer → TimeMixer — each improving MSE by 1–3% and often swapping ranks across datasets. That’s what “saturated” means: lots of strong models, no clear king.

2. The idea: forecasting is sequential, so use online learning
#

Here’s the under-used observation. LTSF evaluation is a sequence: you forecast a window, then time advances and you observe what actually happened, then you forecast the next window. That’s precisely the classic setting of prediction with expert advice / online aggregation: combine several “experts” (here, the trained models) and adapt their weights from the losses you’ve seen so far, with no distributional assumptions.

Yet LTSF papers almost always report a single model. We ask: what if we just combine the existing strong models, online?

3. The method (and what “leakage-free” means)
#

Our pool = {PatchTST, iTransformer, TimesNet, TimeMixer, DLinear}, each trained with the standard library’s official settings. For each forecast window \(t\), each model \(k\) gives a prediction \(\hat f_{k,t}\). We pick non-negative weights summing to 1 that would have minimized error on the past windows, and forecast a weighted average:

$$ w_t=\arg\min_{w\ge 0,\ \mathbf{1}^\top w=1}\ \sum_{sThe crucial detail for newcomers: the weights for window \(t\) use only windows \(s — information available before we forecast \(t\). That’s leakage-free: no peeking at the future or the answer. The first few windows just use a uniform average as warm-up. This program is a tiny linear program in \(K\) weights and has no learning rate or temperature to tune — a real practical advantage over the classic exponential aggregators.

4. Results: beating the best single model on every cell
#

Test MSE, online-convex ensemble vs. the best single model in the pool (lower is better):

Datasetbest singleonline ensemblereductionsig.\(^\dagger\)
ETTh1 (avg)0.4390.430−2.1%
ETTh2 (avg)0.3750.364−2.9%
ETTm1 (avg)0.3820.373−2.4%
ETTm2 (avg)0.2780.274−1.6%
Exchange-7200.7490.482−36%

\(^\dagger\)Diebold–Mariano test, one-sided, \(p<0.05\). Across all 20 (dataset, horizon) cells, the convex/ridge ensemble beats the best single model on 20/20, statistically significant on 15/20 (all ETTm cells, most ETTh; Exchange is noisier — fewer windows — but still wins every horizon). Gains are small where the benchmark is tight (ETT, 1–6%) and large where the models genuinely disagree (Exchange, up to 36%).

5. Frontier vs. frontier: parameter-free wins
#

We didn’t stop at “our thing beats the single model.” We compared against the established online aggregators — exponentially weighted average (EWA), polynomial weighting (MLpol), and Bernstein Online Aggregation (BOA). Number of cells (of 20) each beats the best single model:

simple avgEWAMLpolBOAconvex (ours)ridge (ours)
201612122020

The sophisticated regret-based methods (MLpol, BOA) are actually less robust here — they need a well-chosen learning rate and stumble on the high-variance Exchange horizons — while our parameter-free convex/ridge combiner wins everywhere. Lesson: online aggregation is the lever; a simple non-negative convex program is a strong, tuning-free way to pull it.

6. The same trick generalizes (electricity prices)
#

This isn’t ETT-specific. The identical leakage-free online combiner, applied to the electricity-price-forecasting benchmark’s own models, beats published N-BEATSx on all 5 markets (mean −2.6% MAE, \(p\le0.007\)) — details in the companion post. Two very different domains, one tuning-free method: that’s the sign of a real lever, not a dataset artifact.

7. War stories (the infra was the hard part)
#

The science was clean; the plumbing taught us things:

  • Don’t route giant tensors through a tiny box. Our control node has ~1GB RAM. ECL/Traffic predictions are multi-gigabyte per cell (321/862 channels). Trying to centralize them broke every transfer with “Broken pipe” — the 1GB box couldn’t buffer them. The fix: train and analyze each dataset on the same GPU box (lots of RAM/disk), and only ship back a tiny results file. Never move the big arrays.
  • Right-size the disk. Standard rented GPUs (32GB disk) can’t even store a full ECL/Traffic ensemble. We rented a dedicated 150GB node (at \(\le\$0.2\)/GB-month) specifically to finish the high-dimensional datasets.
  • Memory-bounded analysis. Even reading those preds needs care: we stream them in small blocks (mmap), so the ensemble math runs in <500MB regardless of channel count.

8. Takeaways
#

  • On a saturated benchmark, a leakage-free online ensemble of existing strong models beats the best single one — significantly and everywhere we tested — for almost no extra cost.
  • Parameter-free (convex/ridge) beats the tunable classics (EWA/MLpol/BOA) in robustness.
  • The win transfers across domains (LTSF ↔ electricity prices), suggesting online aggregation is an under-used, general lever.
  • And: respect your infrastructure limits — RAM and disk shape what’s even possible.

In-progress research log. ETT + Exchange numbers are from our own runs under the leakage-free protocol above, read straight from disk; ECL/Traffic are finishing on the dedicated big-disk node and will be added.