Learned Forecast Skeletons for Frozen Time-Series Foundation Models

Text-to-image models misspell words because their text encoders keep what a word means and throw away what it looks like. The fix that worked, called glyph conditioning, feeds the image generator a plain black-and-white rendering of the target text as an extra input at every denoising step. The generator then only has to decide how the letters should look, while the blueprint fixes what shape they must have. Time-series foundation models have a similar problem. These are large sequence models, pretrained on many domains, that forecast unseen series without further training. They see only the recent past, forecast step by step or patch by patch, and drift away from the series' underlying seasonal shape over long horizons and on noisy, high-frequency data. We propose to give such a frozen model a forecast skeleton: a smooth, scale-free sketch of the likely future shape, supplied as an extra input channel. A small module builds this sketch from similar series in a library drawn from other domains.

The central claim is about how to train that small module. At deployment the target domain is new, so every retrieved neighbor comes from a different domain. If the module is trained the usual way, with neighbors retrieved from the same domain, it learns to copy near-duplicates. That habit breaks under a domain shift. We therefore train it in episodes, the core idea of meta-learning for domain generalization: each training step holds out one source domain, restricts the retrieval library to the remaining domains, and scores the frozen model's forecast on the held-out domain. The hypothesis is that this match between training and deployment, rather than the template channel alone, is what makes skeleton conditioning transfer across domains.

Core Claim and Relation to Prior Work

Three kinds of prior work shape the design. The first supplies the mechanism. GlyphControl feeds rendered glyph layouts through a trainable side branch while the base generator stays frozen. GC-DDPM and DiffCJK concatenate the glyph image with the noisy input at every step so that geometry cannot drift. From these we take three design choices: the template is canonical rather than stylized, it is present at every forecast position rather than only at the start, and the large model stays frozen.

The second kind of work motivates the proposal by documenting the failure. GIFT-Eval reports that foundation models lose to PatchTST and statistical baselines on high-frequency noisy data and over long horizons, and that step-by-step decoders accumulate error. These are exactly the regimes where a fixed shape anchor should help most.

The third kind is the closest existing work. TARPA averages the retention curves of the five most similar videos and feeds the average as an extra channel into a frozen foundation model. It reports a Spearman rank correlation of 0.9560 against 0.9375 for CatBoost. However, it works on 85 videos from a single creator, and its template is a fixed average that is never trained. Our proposal differs in three ways: the skeleton is produced by a trained module, the library is deliberately taken from other domains, and training follows the leave-one-domain-out condition that deployment imposes.

Latent Temporal Generalization (arXiv version) is the closest domain-generalization work in forecasting. It splits series into trend and seasonal parts and learns latent factors that separate what domains share from what is specific to each, using a conditional variational autoencoder. It uses small forecasters and fixed regularization penalties, with no foundation model and no episodic training. We keep its insight that decomposed structure transfers, but express the structure as an explicit curve the frozen model can read, not as a latent vector.

On the meta-learning side, MASF and shape-aware meta-learning for prostate MRI pair episodic domain rotation with structural or shape priors in images. MASF reports 75.21% on PACS against 71.70% for simple pooling of all source domains. These papers also constrain the design. MASF reports costly second-order gradients and fragile hyperparameters, so we deliberately use a first-order episodic scheme: the only thing that changes across episodes is which domains the library may contain. There is no inner-loop update.

Method

Let $F$ be a frozen foundation model that accepts extra input channels. Our primary choice is Chronos-2, whose group attention reads covariates without forecasting from them directly. That avoids the "picky eater" failure TARPA describes, where a foundation model is disrupted by inputs that are not raw trajectories. MOMENT serves as a second backbone, with the skeleton passed as an extra channel.

Each library entry is a context window paired with its continuation. Every entry is canonicalized: it is normalized to zero mean and unit variance, then low-pass filtered. This is the time-series analogue of a binary glyph, keeping shape and discarding amplitude and noise.

For a target context $x$, we retrieve the $k$ nearest library windows by shape distance. The skeleton module $g_\theta$ is a small network, well under a million parameters. It aligns the neighbors' continuations to $x$ by rescaling, phase shift and mild time-warping, then blends them together with an STL seasonal profile of $x$. The output is a skeleton $\tau$ covering the context plus the forecast horizon. Training minimizes

$$\min_\theta \; \mathbb{E}_{d \sim \mathcal{D}} \; \mathbb{E}_{(x,y) \sim d} \; \ell\big(F(x, g_\theta(x; \mathcal{L}_{\mathcal{D}\setminus d})),\, y\big),$$

where $\mathcal{D}$ is the set of source domains, $d$ is the held-out domain in an episode, $\mathcal{L}_{\mathcal{D}\setminus d}$ is the canonicalized library built only from the other domains, $y$ is the true future, and $\ell$ is the backbone's probabilistic loss. Gradients pass through the frozen $F$ into $\theta$ only.

Because the STL profile is an input to $g_\theta$, the module can learn to fall back on classical decomposition when cross-domain neighbors are useless.

Research Program

The program asks three questions. First, does skeleton conditioning improve a frozen foundation model on domains held out entirely? Second, is the improvement caused by episodic library restriction, rather than by the extra channel or by retrieval itself? Third, does the gain concentrate where GIFT-Eval says foundation models fail, namely long horizons and high-frequency noisy series?

We also analyze what the skeletons contain. For each target we measure how much weight $g_\theta$ places on cross-domain neighbors versus the STL profile, and whether the neighbors it chooses come from domains with similar sampling frequency. If the module simply collapses to the STL profile, the cross-domain part of the idea has failed, and we will report that.

Baselines

The naive reference points are seasonal-naive and Auto_Theta forecasts. GIFT-Eval found Auto_Theta competitive on noisy data. The frozen backbone with no extra channel is the do-nothing version of our system.

The strongest existing method is zero-shot Chronos-2. It reports a 90.7% average win rate on fev-bench over TiRex and Moirai-2.0. For the per-domain datasets we use, there is no reported result found on Lacuna for Chronos-2 on GIFT-Eval held-out domains or on the Latent Temporal Generalization datasets, so we will run it ourselves. We will use identical context lengths and horizons, and report only datasets outside its declared pretraining data, which GIFT-Eval's non-leaking splits make possible.

PatchTST is the strongest supervised competitor according to GIFT-Eval, which found it the most dominant model on average (no numeric cell reported on Lacuna). It will be trained on the source domains under the same leave-one-domain-out protocol and tuned with the same number of trials as our module.

The TARPA-style baseline feeds the untrained average of the $k$ nearest canonicalized continuations to the same frozen backbone through the same channel. It isolates whether training the skeleton module matters at all.

Latent Temporal Generalization with its WaveNet and DLinear forecasters is the domain-generalization baseline on Favorita-cat, Web-traffic, Stock-volume and Power-cons. It is the natural baseline there because it was built for that setting; its paper reports the lowest $Q(0.5)$ and $Q(\text{mean})$ on Favorita-cat with WaveNet, but no numeric value is found on Lacuna. We will rerun it using its released protocol.

The key ablation trains exactly the same skeleton module by ordinary pooled training. It uses the same data, steps and tuning budget, but retrieves from a library that includes the query's own domain, with the query series itself excluded. Beating this ablation is the core claim.

We count the method as a success only if two conditions hold. It must reduce geometric-mean normalized CRPS and MASE over held-out domains relative to both zero-shot Chronos-2 and the TARPA-style baseline, with the reduction statistically reliable under paired tests over series and three seeds. And the gap to the pooled-training ablation must widen, not shrink, on the held-out domain.

Why not just…?

Why not just use the foundation model zero-shot, or a larger one? GIFT-Eval shows scale does not close the gap on noisy high-frequency and long-horizon series. SEMPO reports a 23.1% zero-shot MSE reduction over far larger models, including Chronos, with a 6.5M-parameter model, which suggests structure can matter more than size. Our horizon-stratified comparison against the zero-shot backbone answers this directly for our setting.

Why not just pass the TARPA nearest-neighbor average as a covariate? This is a real threat. On strongly seasonal domains with close cross-domain neighbors it may match us. The leave-one-domain-out comparison settles it. We expect the untrained average to hurt when neighbors are misaligned in phase or scale, because nothing corrects the misalignment. If it matches us everywhere, the trained module is unnecessary and we will say so.

Why not just use an STL seasonal profile as the template? This is also a real threat, and it is why the STL profile is an input to our module. We will report an STL-only channel baseline per dataset. Our module can only justify its cost where cross-domain neighbors add shape information that the target's own short context cannot provide, for example cold-start series with less than one full period of history.

Why not just fine-tune the backbone with episodic training, or run full second-order meta-learning? Fine-tuning a large pretrained model on a few source domains risks overwriting its general knowledge. Second-order meta-learning carries the cost and instability MASF documents. We include a parameter-efficient fine-tuned Chronos-2, trained with the same episodic protocol and the same compute budget, as a comparison. If it wins, the argument for keeping the backbone frozen and training only the template module fails.

Why not just train the skeleton module normally? That is the pooled-training ablation above. SPUME shows that ordinary training latches onto shortcuts that break under shift, lifting Waterbirds worst-group accuracy from 66.4% to 85.9%. Same-domain near-duplicates are the retrieval version of such a shortcut. If pooled training performs as well on held-out domains, the episodic claim is false.

Decisive Checks

We run leave-one-domain-out evaluation on GIFT-Eval's seven domains, reporting normalized CRPS and MASE by frequency and horizon. The four Latent Temporal Generalization datasets, where domains are stores, language projects, stock markets and power networks, provide a second domain-generalization test. fev-bench is used as a secondary check that the extra channel does not hurt the backbone where it is already strong.

Four outcomes would decide the question. First, the episodic module must beat the pooled-training ablation on held-out domains while the two tie on in-domain test data. That pattern is the signature of a train/deployment mismatch, and is the heart of the claim. Second, the gains over zero-shot Chronos-2 must grow with horizon length and be largest on high-frequency domains. Third, replacing skeletons at test time with skeletons from randomly chosen other series must erase the gains, which confirms the backbone actually reads the channel. Fourth, removing the low-pass canonicalization must weaken cross-domain transfer, which would support the glyph analogy that shape without style is what transfers.