Skip to main content

​
What is Zero-Shot Forecasting?

Zero-shot forecasting allows you to make predictions on time-series data without fine-tuning the model on your specific dataset. Samay’s pre-trained foundation models (LPTM, MOMENT, Chronos, Moirai, etc.) can generate forecasts immediately using their learned representations from large-scale pre-training. This approach is ideal when:
  • You have limited training data
  • You need quick predictions without training time
  • You want to test a model’s capabilities before committing to fine-tuning
  • Your data shares similar patterns with the pre-training corpus

​
Step-by-Step Workflow

Here’s how to perform zero-shot forecasting with any Samay model:
1

Import the model

First, import and initialize your chosen model:
from samay.model import LPTMModel  # or MomentModel, ChronosModel, etc.

config = {
    "task_name": "forecasting2",
    "forecast_horizon": 192,
    "seq_len": 1112,
    "head_dropout": 0,
    "weight_decay": 0,
    "max_patch": 16,
    "freeze_encoder": False,
    "freeze_embedder": True,
    "freeze_head": True,
    "freeze_segment": True,
}
model = LPTMModel(config)
2

Prepare your dataset

Create a dataset instance for your time-series data:
from samay.dataset import LPTMDataset

val_dataset = LPTMDataset(
    name="ett",
    datetime_col="date",
    path="../data/data/ETTh1.csv",
    horizon=192,
    task_name="forecasting2",
    seq_len=1112,
)
3

Run zero-shot evaluation

Evaluate the model directly without training:
avg_loss, trues, preds, histories = model.evaluate(
    val_dataset, task_name="forecasting2"
)
print(f"Average Loss: {avg_loss}")
4

Visualize results

Plot the predictions against ground truth:
import matplotlib.pyplot as plt
import numpy as np

# Convert to numpy arrays
trues = np.array(trues)
preds = np.array(preds)
histories = np.array(histories)

# Pick a random sample
channel_idx = np.random.randint(0, 7)
time_index = np.random.randint(0, trues.shape[0])

history = histories[time_index, channel_idx, :]
true = trues[time_index, channel_idx, :]
pred = preds[time_index, channel_idx, :]

plt.figure(figsize=(12, 4))
plt.plot(range(len(history)), history, label="History", c="darkblue")

offset = len(history)
plt.plot(
    range(offset, offset + len(true)),
    true,
    label="Ground Truth",
    color="darkblue",
    linestyle="--",
    alpha=0.5,
)
plt.plot(
    range(offset, offset + len(pred)),
    pred,
    label="Forecast",
    color="red",
    linestyle="--",
)
plt.legend()
plt.show()

​
Real Examples

​
LPTM Zero-Shot Forecasting

From lptm_zero.ipynb, LPTM achieves strong zero-shot performance on the ETTh1 dataset: 
from samay.model import LPTMModel
from samay.dataset import LPTMDataset

# Initialize model
config = {
    "task_name": "forecasting2",
    "forecast_horizon": 192,
    "seq_len": 1112,
}
model = LPTMModel(config)

# Load dataset
val_dataset = LPTMDataset(
    name="ett",
    datetime_col="date",
    path="data/ETTh1.csv",
    horizon=192,
    task_name="forecasting2",
    seq_len=1112,
)

# Zero-shot evaluation
avg_loss, trues, preds, histories = model.evaluate(
    val_dataset, task_name="forecasting2"
)
# Results: avg_loss ≈ 0.376

​
Chronos Zero-Shot Forecasting

Chronos provides probabilistic forecasts with quantile predictions: 
from samay.model import ChronosModel
from samay.dataset import ChronosDataset

repo = "amazon/chronos-t5-small"
chronos_model = ChronosModel(repo=repo)

val_dataset = ChronosDataset(
    name="ett",
    datetime_col="date",
    path="data/ETTh1.csv",
    mode="test",
    batch_size=8
)

# Visualize zero-shot forecasting with uncertainty
chronos_model.plot(
    val_dataset,
    horizon_len=64,
    quantile_levels=[0.1, 0.5, 0.9]
)

​
Tips for Best Results

Use longer context windows for capturing long-term patterns. Most models support 512-1024 timesteps.
Ensure your data is properly scaled. Most Samay datasets handle normalization automatically.
If using models like Moirai, specify the frequency parameter (freq='h' for hourly, freq='d' for daily, etc.).
Different architectures excel at different patterns. Test LPTM, MOMENT, Chronos, and Moirai to find the best fit.
For critical applications, combine predictions from multiple zero-shot models.

​
Next Steps

  • Poor zero-shot results? Try fine-tuning on your specific dataset
  • Need uncertainty estimates? Use probabilistic models like Chronos or Moirai
  • Multivariate forecasting? LPTM and MOMENT support multiple time-series channels
Ready to improve performance? Check out the Fine-Tuning Guide.

Build docs developers (and LLMs) love

Get started for freeTalk to us