PatchTST — Patch-based Time Series Transformer

Multivariate forecasting · Fine-tunable · HuggingFace-native

---

Overview

• Model ID: patchtst-forecasting
• Architecture: PatchTSTForPrediction
• Framework: HuggingFace Transformers
• Backend: forecasting specialist (specialists-ts-cu121)
• specialist_type: forecasting
• Hardware: GPU (small footprint, ~512 MB VRAM); also runs on CPU (slower)
• Training time: 5–60 minutes depending on dataset size

PatchTST divides a time series into patches (like ViT patches for images) and processes them with a Transformer encoder, reaching strong results on standard benchmarks with fewer parameters than channel-mixing models.

---

When to Use

✅ Perfect for:

• Multivariate forecasting (multiple correlated series)
• ETT-style benchmarks (energy, weather, traffic)
• When you have enough data to fine-tune
• Point forecasts with a fixed horizon

❌ Not ideal for:

• Zero-shot forecasting — use TimesFM 2.5
• Very short series
• Probabilistic forecasts / quantiles — use TimesFM
• Irregular timestamps — use Prophet

---

Training

Fine-tune PatchTST on your own time series.

from colabhive import ColabHive

client = ColabHive(api_key="YOUR_KEY", account_id="YOUR_ACCOUNT")

# CSV time series: a value column (and optional extra channels), ordered in time
dataset = client.datasets.upload(name="my-timeseries", file="./train.csv")

job = client.training.create(
    model="patchtst-forecasting",
    dataset_id=dataset.id,
    hyperparameters={
        "context_length": 512,
        "prediction_length": 96,
        "patch_length": 16,
        "stride": 8,
        "d_model": 128,
        "num_attention_heads": 16,
        "num_hidden_layers": 3,
        "learning_rate": 1e-4,
        "num_train_epochs": 20,
        "per_device_train_batch_size": 32,
    },
)
job.wait()
print(job.get_metrics())

The exact accepted hyperparameters come from the template's schema: GET /api/builder/v1/training/model-configs/{model_config_id}/schema.

Register for inference

endpoint = client.training.register_for_inference(
    run_id=job.id,
    name="patchtst-energy-forecast",
    description="PatchTST multivariate forecaster",
    visibility="account",
)

---

Inference

# Univariate: list of floats. Multivariate: list of channel lists.
result = client.endpoints.infer(
    endpoint_id=endpoint.endpoint_id,
    input_data={
        "values": [1.2, 1.5, 1.3, 1.8, 2.1, 2.4, 2.2, 2.6],  # historical points
        "horizon": 96,   # steps to predict
    },
)
print(result)

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/{ENDPOINT_ID}/infer" \
  -H "X-Account-ID: YOUR_ACCOUNT_ID" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"values": [1.2, 1.5, 1.3, 1.8, 2.1, 2.4], "horizon": 24}}'

Multivariate input — pass one list per channel under values.

---

Hyperparameter Guide

Parameter	Effect	Typical Range
context_length	How much history the model sees	96–2048
prediction_length	Forecast horizon	24–720
patch_length	Patch size (like kernel size)	8–64
stride	Overlap between patches	patch_length / 2
d_model	Transformer hidden dim	64–512
num_hidden_layers	Transformer depth	2–6
num_attention_heads	Attention heads	4–16
dropout	Regularization	0.1–0.3

Typical configs by dataset size:

Dataset size	context_length	d_model	layers	epochs
Small (< 1k windows)	96	64	2	50
Medium (1k–10k)	336	128	3	30
Large (> 10k)	512	128	3	20

For the original benchmark numbers, see the PatchTST paper.

---

Related

• BiTCN Forecasting — TCN alternative, faster training
• TimesFM 2.5 — zero-shot foundation model
• Classical Forecasting — Prophet, ARIMA