Walk-Forward Validation Calculator for Machine Learning

1. Dataset and time settings

Dataset source

Total observations

CSV row count overrides this value.

CSV dataset

Recognised headers include date, actual, predicted, target, forecast, and yhat.

Saved configuration JSON

Loads previously exported form settings in your browser.

Date column name

Actual or target column

Prediction column

Missing-value handling

Duplicate timestamp handling

Sort CSV timestamps chronologically

Dataset start date

Dataset end date

Time frequency

Custom interval days

Date display format

Prediction task

Decimal precision

2. Validation windows

Walk-forward strategy

Initial training size

Initial-size unit

Rolling-window size

Used by rolling and sliding strategies.

Test-window size

Forecast horizon

Step size

Train-test gap

Embargo periods

Maximum folds

Enter zero to generate every valid fold.

Skip test-start indices

Use this to exclude holidays, outages, or unavailable forecast origins.

Allow overlapping test windows

3. Retraining and forecasting setup

Retraining policy

Periodic retraining frequency

Average training seconds per fold

Average prediction seconds per fold

Forecasting method

Model

Primary metric

Seasonal period

Positive class

Probability threshold

Concept drift is expected

Features are leakage-sensitive

4. Time-series preprocessing

Lag-feature count

Rolling mean window

Rolling deviation window

Differencing order

Seasonal differencing order

Scaling method

Apply log transformation

Fit preprocessing on training folds only

Inverse-transform predictions

5. Actual and predicted values

Enter a complete dataset series or sequential fold predictions. Classification probabilities must remain between zero and one.

Actual values or labels

Predicted values, labels, or probabilities

6. Scenario comparison

Compare a second validation configuration

Scenario B strategy

Scenario B initial size

Scenario B rolling size

Scenario B test size

Scenario B step size

Scenario B gap

Formula used

Folds = floor((N − I − H − G) ÷ S) + 1

N is the observation count. I is the initial training size. H is the test window, G is the effective gap, and S is the step size.

For rolling validation, each training start moves forward. Expanding validation keeps the original training start.

How to use this calculator

Enter the observation count or upload a CSV file.
Select an expanding, rolling, sliding, anchored, or prequential strategy.
Set training, testing, horizon, step, gap, and embargo values.
Choose the model, retraining policy, preprocessing, and primary metric.
Optionally provide actual and predicted values for metric calculations.
Review folds, leakage notices, charts, recommendations, and exports.

Example validation schedule

Fold	Training indices	Testing indices	Purpose
1	0–59	60–71	Initial twelve-period forecast
2	0–71	72–83	Expanded training history
3	0–83	84–95	Later out-of-sample evaluation

Frequently asked questions

What is walk-forward validation?

It evaluates time-ordered models using repeated training and future testing windows. The method imitates real forecasting more closely than random splitting.

When should I use an expanding window?

Use it when older observations remain informative. Every new fold adds more historical training data.

When should I use a rolling window?

Use it when recent behaviour matters more. It limits training history and can respond better to drift.

How should I choose the initial training size?

Include enough observations for stable fitting and seasonal coverage. Complex models usually require longer initial windows.

What does the step size control?

It controls how far the validation origin advances. Smaller steps create more folds and possible overlap.

Why add a train-test gap?

A gap helps separate training features from future labels. It can reduce leakage with delayed or overlapping targets.

What is an embargo period?

An embargo excludes nearby observations around validation boundaries. It is useful when target information overlaps through time.

Can this calculator train an ARIMA or LSTM model?

It plans validation and evaluates supplied predictions. Model training should run in your preferred machine-learning environment.

How are per-fold metrics assigned?

Complete dataset predictions use test indices directly. Shorter series are assigned sequentially when every fold horizon is present.