How to Run a Walk-Forward Test on a Trading Strategy

A backtest tuned on ten years of data and then scored on the same ten years has measured one thing, how well the settings fit the past. It says nothing about next year. Walk-forward testing asks what the rules would have done on data they had never seen, and it asks repeatedly across the history, one unseen stretch at a time, so a single lucky year cannot carry the verdict.

How do you split the history?

Settle the lengths before you look at a single result: a tuning window, where the settings are chosen, and a test window, where they are used. The choice depends on how often the strategy trades. Each tuning window needs enough trades to pick settings on, and each test window needs enough to judge.

With ten years of daily data and a strategy that trades a few dozen times a year, four years to tune and one year to test is a reasonable start.

How does one pass work?

Tune on years 1 to 4. Find the moving-average length, the stop distance, whatever the strategy’s settings are, that performed best in those years. Write them down.

Then run those exact settings on year 5. Don’t touch them. If year 5 looks bad and a small tweak would fix it, the tweak is the thing the test is designed to catch.

Costs go in from this first pass. A hypothetical strategy trading 50 times a year with $40 of commission and slippage per round trip gives up $2,000 a year before it earns anything, and a tuning run without costs will happily choose settings that trade too often to survive them.

How do you roll it forward?

Slide both windows along by one test length and do it again.

Each pass may choose different settings. That’s fine. It is how the strategy would really be run, retuned on recent data at regular intervals. What you are testing is the whole process, the rules plus the way you choose their settings, and that is the thing you would actually be trading with real money.

This version rolls, dropping the oldest year each time. An anchored version keeps year 1 as the start and lets the tuning window grow. Both are valid. Pick one before you start.

What do you do with the results?

Stitch the test years together into one record. Leave every tuning result out. That record is the one to judge: its return, its drawdown, its longest losing streak.

Then compare it with the tuning windows. Suppose, hypothetically, the tuning windows averaged 20% a year and the stitched test years averaged 8%. The test kept 8 / 20 = 40% of the tuned performance. Some drop is normal, since tuned results always flatter. A test record that falls to nothing, or turns negative, says the settings were fitted to noise in each window and the strategy has no edge you can find this way.

Look at the settings each pass chose, as well. If the best moving-average length drifts gently, say from 45 to 50 days across the passes, the strategy has a stable region worth trusting. If it jumps from 20 to 90 and back, each window’s winner was a fluke of that window. That instability is a finding in its own right.

Watch the spread of test years too. One great year and five flat ones is a different strategy from six modest ones, even with the same average, and the losing streak calculator will tell you how deep a run of losses the stitched record implies for your position size.

Where does it stop protecting you?

It can’t show you a market regime missing from all ten years. It can’t catch a data error that runs through the whole history, such as a list of stocks that leaves out companies that were delisted. And six test years is still a small sample of years.

The case for trusting only out-of-sample numbers is argued in out-of-sample testing is the only part that counts. For building rules a test can actually run, the course Rules First, Money Later starts from scratch, and its lesson on backtesting biases and costs covers the errors a walk-forward test won’t catch on its own.