Appearance

Personalize your experience

Mode

Accent Color

Direction

Walk-Forward Analysis

Backtesting

Walk-forward analysis repeats what deployment does: choose parameters on the past, trade them on a future they have not seen, move forward and choose again. Only the out-of-sample windows count.

The windows

Mode	In-sample window	Use it when
Rolling	a number of bars just before each out-of-sample	markets change and old is less relevant
Anchored	from the first bar up to each out-of-sample (it grows)	the edge is structural and more helps

Each : every grid combination is backtested on the in-sample bars, the best by your is kept, and it is run on the next out-of-sample bars, starting flat. The next window starts where that one ended. With 10,000 bars, 3,000 in-sample and 1,000 out-of-sample give seven windows.

Reading the result

Measure	Meaning	Healthy
Out-of-sample return per	what trading would have earned	mostly positive, no single carrying the total
Profitable OOS windows	share of windows with a profit	60% or more
Efficiency	OOS return per bar / IS return per bar	about 0.5 or higher; near 0 means the in-sample do not carry over
Parameter changes	how often the chosen set differs from the previous	few; constant switching means the optimum is noise

Caution:Do not tune the walk-forward itself ( sizes, , grid) until it looks good: that is another search over the same , and the out-of-sample windows stop being out of sample.

Choosing window sizes

The in-sample window should hold enough for the to mean something: aim for 30 or more per .
The out-of-sample is how long you would trade between re-optimizations. A quarter of the in-sample length is a common start.
Indicator warm-up eats the start of each out-of-sample on the server (it starts flat). Very short OOS windows with long indicators trade little.

API and SDK

POST /backtest/engine/walk-forward

{ "data": { "...": "..." }, "strategy": { "...": "..." },
  "walkForward": { "inSample": 3000, "outOfSample": 1000, "mode": "rolling", "objective": "sharpe",
                   "grid": { "fast": [5, 10, 20], "slow": [40, 80] } } }

quatick.optimize.walk_forward

wf = optimize.walk_forward(strategy, bars, {"fast": [5, 10, 20], "slow": [40, 80]},
                          in_sample=3000, out_of_sample=1000, mode="anchored", n_jobs=8)
wf["oos_stats"]["sharpe"], wf["efficiency"], wf["profitable_windows_pct"]
wf["oos_equity"]   # stitched out-of-sample equity (the SDK warms indicators up on in-sample history)

Appearance

Walk-Forward Analysis

The windows

Reading the result

Choosing window sizes

API and SDK

See also

Related articles

Appearance

The windows

Reading the result

Choosing window sizes

API and SDK

See also

Related articles

Related to Walk-Forward Analysis