Skip to main content

​
Preparing time series

Forecast quality is tied to how you curate the source data. The SDK enforces some guard rails, but high-signal forecasts start with clean inputs.

​
Rule of thumb

  • Provide at least 30 observations for robust forecasts.
  • Keep series consistent: no NaNs, no None, and a single measurement per interval.
  • Detrend and normalise when you mix heterogeneous scales.

​
Filling gaps

import pandas as pd

raw = pd.Series(
    [110, None, 118, 120, None, 131],
    index=pd.date_range("2024-01-01", periods=6, freq="D"),
)
clean = raw.interpolate(method="time").fillna(method="ffill")
series = clean.astype(float).tolist()
The SDK converts numpy arrays to lists during serialisation, so broadcasting to floats ahead of the request keeps payloads lean.

​
Scaling and clipping

import numpy as np
from simulacrum import Simulacrum

client = Simulacrum(api_key="sim-key_id-secret")
window = np.asarray(series, dtype=float)
window = np.clip(window, a_min=0, a_max=None)
window = (window - window.mean()) / window.std()

forecast = client.forecast(series=window, horizon=10, model="tempo")
Log transforms can also stabilise variance for multiplicative series before you forecast.

​
Feature flags

Store the context you use to shape the series alongside the request so you can replay it later if an issue surfaces. 
metadata = {
    "segment": "north-america",
    "transforms": ["interpolate", "zscore"],
    "model": "tempo"
}
Log both the metadata and the response body in your data warehouse to power audit trails.