Time Series

Overview

Time series models predict future values based on past observations that are ordered in time (for example: daily sales, hourly traffic, or weekly revenue). Unlike standard ML models, time series models are sensitive to time spacing. This means the model must know what one row of data represents (day, hour, week, etc.) and how far back in time it should look when learning patterns. This section explains the key configuration options specific to time series models and how to use them correctly.

Required Columns

Timestamp Column

Required: Yes

The name of the column that contains the date or timestamp for each observation.

• Used to order the data in time
• Must contain valid date or datetime values
• Each row should represent a single point in time

Target Column

Required: Yes

The Target Column is the numeric value the model will learn to predict. This column represents the time-dependent quantity you want to forecast.

Extra Regressors

Required: No

Additional feature columns that may help explain changes in the target value. These features must be known at prediction time or remain constant in the future.

Time Configuration

Frequency

Defines what one row in your dataset represents.

Examples:

• D → one row per day
• H → one row per hour
• W → one row per week
• M → one row per month

⚠️ Important: Frequency must match the actual spacing of your data.
If your data is daily but Frequency is set to W, the model will treat each row as one week apart, leading to incorrect lags, rolling features, and forecasts.

If it is not provided, it will be inferred from the timestamps in your data.

Forecast Steps

Determines how many future time periods the model will predict. The unit of each step depends on the selected Frequency.

Examples:

• Forecast Steps = 7 with Frequency = D → forecast the next 7 days
• Forecast Steps = 4 with Frequency = W → forecast the next 4 weeks

Fill Method

How missing timestamps are handled before training.

Available options:

• interpolate — linearly interpolate missing values
• ffill — forward fill using the last known value
• bfill — backward fill using the next known value

Choose a method that aligns with how gaps should reasonably be filled in your data.

Add Lags (Days)

Allows the model to use past values of the target as features. Each value represents how many time steps (rows) back to look, not calendar days.

Examples:

• With Frequency = D and Add Lags (Days) = [1, 7]
  • 1 → yesterday
  • 7 → 7 days ago
• With Frequency = W and Add Lags (Days) = [1, 7]
  • 1 → last week
  • 7 → 7 weeks ago

Lag features help capture short-term patterns and repeating behavior.

Add Rolling Mean

Adds rolling average features based on recent past values. Each value represents the window size in time steps (rows).

Examples:

• With Frequency = D and Add Rolling Mean = [7, 30]
  • 7 → average of the last 7 days
  • 30 → average of the last 30 days
• With Frequency = W and Add Rolling Mean = [4]
  • 4 → average of the last 4 weeks

Rolling means smooth short-term noise and capture the recent overall level of the series.