How to Clean and Resample Time Series¶

This guide shows you how to prepare raw data for a forecasting pipeline: validating dtypes and value ranges, then changing the series frequency with Downsampler and Upsampler.

Prerequisites¶

Familiarity with polars DataFrames and the "time" column convention (Core Concepts)

Try it interactively¶

How to Clean Time Series Data

End-to-end data cleaning pipeline combining SimpleTimeImputer and SeasonalImputer for missing values with OutlierThresholdHandler for anomaly clipping.

View · Open in marimo
How to Resample Time Series

Demonstrate Downsampler and Upsampler for changing time series frequency, including multivariate support, boundary settings, and round-trip information loss.

View · Open in marimo

Validate Column Types¶

Yohou expects the "time" column to be Datetime or Date and all value columns to be numeric. If your data arrives with string columns, cast them before fitting:

import polars as pl

df = df.with_columns(
    pl.col("time").str.to_datetime(),
    pl.col("value").cast(pl.Float64),
)

Run df.dtypes to verify the schema before proceeding.

Clean Implausible Values¶

Sentinel values and out-of-range readings can bias model training. Replace them with null so that downstream imputation or aggregation handles them consistently:

df = df.with_columns(
    pl.when(pl.col("value") < 0)
    .then(None)
    .otherwise(pl.col("value"))
    .alias("value")
)

If the valid range is known, clip is a simpler alternative:

df = df.with_columns(pl.col("temperature").clip(-50, 60))

Downsample with Downsampler¶

Downsampler reduces frequency by aggregating within each period. Use it when the data arrives at a higher frequency than you need for forecasting, or when high-frequency noise obscures the signal:

from yohou.preprocessing import Downsampler

downsampler = Downsampler(interval="1mo", aggregation="mean")
df_monthly = downsampler.fit_transform(df)

The aggregation parameter accepts "mean", "sum", "min", "max", "first", "last", or "median". Choose the one that matches the quantity: "sum" for totals (revenue, rainfall), "mean" for rates and averages (temperature, speed), "last" for snapshot values (inventory, account balance).

For finer control over bin boundaries, pass closed and label:

downsampler = Downsampler(
    interval="1w",
    aggregation="sum",
    closed="right",
    label="right",
)

Upsample with Upsampler¶

Upsampler increases frequency by creating new timestamps and filling in values. Use it when your model needs a finer resolution than the source data provides:

from yohou.preprocessing import Upsampler

upsampler = Upsampler(interval="1d", interpolation="linear")
df_daily = upsampler.fit_transform(df)

The interpolation parameter controls how gaps are filled:

Method	Behaviour	Good for
`"linear"`	Interpolates between known points (default)	Smooth, continuous quantities (temperature, price)
`"forward"`	Carries the last known value forward	Step quantities that hold until updated (inventory, status)
`"backward"`	Fills from the next known value	Pre-announced values (scheduled rates, published targets)
`"nearest"`	Uses the closest known value in either direction	Sparse data where directionality does not matter

Fill Gaps After Resampling¶

Resampling (or cleaning out implausible values) can leave null gaps that a forecaster cannot consume. Use SimpleTimeImputer to fill them with a time-aware method such as linear interpolation:

from yohou.preprocessing import SimpleTimeImputer

imputer = SimpleTimeImputer(method="linear")
df_filled = imputer.fit_transform(df_monthly)

The method parameter accepts "linear", "forward", "backward", "nearest", or "fill_both". When the gaps follow a repeating cycle, prefer SeasonalImputer, which fills each missing value from the same seasonal position. See Handle Missing Data for the full imputation workflow.