Skip to main content
Track Adaptive moving average
24 / 24
Search the archiveLibrary
Library

2020issue C1132-36

Walk-forward and adaptive averages as two tests of the same trend

Two trend-following implementations were compared under a shared risk constraint of about 10 percent: a walk-forward test path and a Kaufman adaptive moving average path. The recorded net-asset-value paths were close, and a walk-forward test remains out-of-sample only if it is run once.

  • Evaluation here means comparing two implementations of the same trading idea under a shared risk constraint so the result speaks to test design, not to a preferred indicator.
  • A walk-forward test remains an out-of-sample evaluation only if it is run once. Inspecting results, changing a rule, and retesting converts the unused window into another in-sample search.
  • The stated preference for the adaptive moving average was that it does not require periodic retesting and removes the chance of retesting with different in-sample and out-of-sample periods.
  • The closing claim is that when prices trend, trend methods work, and when they do not, none of the methods work. Timing of returns can differ, but long-run results are expected to be similar.
Entries in this reading3 entries
Current trackAdaptive moving average28 readingsTrend following89 readingsWalk-forward analysis95 readings

Two implementations under one constraint

Trend following is a family of entry, exit, and hold rules that stay with a measured direction of price until that direction fails, rather than forecasting a reversal.

This archive evaluation compared two implementations of that idea under a shared risk constraint of about 10 percent: a walk-forward test path and a Kaufman adaptive moving average (KAMA) path.

Walk-forward analysis is a testing procedure that repeatedly optimizes a rule on an in-sample window, then records only the next unused out-of-sample window, and concatenates those unused windows into one evaluation path.

An adaptive moving average is a trend filter that changes its effective lookback as market noise and speed change, so the same rule can slow in choppy periods and speed up when prices persist.

Walk-forward portfolio NAV after 10% volatility targeting

After the four-futures book is scaled to a 10 percent annualized-volatility target, walk-forward net asset value compounds near the 7.2 percent yearly pace Kaufman reported, with a sharp lift around 2015 and another pop in 2018. Point values were read from the published NAV curve, not from a table.
After the four-futures book is scaled to a 10 percent annualized-volatility target, walk-forward net asset value compounds near the 7.2 percent yearly pace Kaufman reported, with a sharp lift around 2015 and another pop in 2018. Point values were read from the published NAV curve, not from a table.E-mini S&P, US 30-year bonds, euro, crude oil · February 1995 to February 2019 · 1995-02-01T00:00:00.000Z to 2019-02-28T00:00:00.000Z

Five-year in-sample window, six-month out-of-sample step, long-only in bonds and the S&P, long and short in the euro and crude. Plotted NAV points are approximate readings from the raster; the 7.2 percent annualized return is the figure stated in the article.

What the recorded paths showed

From March 1993, the walk-forward path returned 7.2 percent annually and the KAMA path returned 7.1 percent annually.

KAMA produced a reward-to-risk ratio of 0.74, slightly higher than the 0.72 recorded for the walk-forward test.

The two net-asset-value paths were described as close overall, with a difference in pattern: KAMA made new highs in the recent market, while later periods were expected to keep the paths tracking closely.

How the speed was chosen

The parameter choice targeted a trend whose speed could vary but averaged about 60 days. That choice was selected from the type of system desired rather than from repeated out-of-sample inspection.

The stated preference for KAMA was that it does not require periodic retesting and removes the chance of retesting with different in-sample and out-of-sample periods.

When trend methods work

The evaluation's closing claim is that when prices trend, trend methods work, and when they do not, none of the methods work. Timing of returns can differ, but long-run results are expected to be similar.

Educational research material, not investment advice. Historical source context does not establish present-day performance.
24 of 24 in the Adaptive moving average track
1994Track finished · Next track: Base buildingCup-and-handle base construction and volume breakout34 readings
All readings on this track · 24 readings
  1. 1991Building variable-length moving averages from partitioned price changes
  2. 1991Variable-length moving average from change dispersion
  3. 1992Constructing volatility-adaptive exponential smoothing
  4. 1995Constructing an adaptive moving average with an efficiency ratio and filter
  5. 1995Two-gate breakout confirmation with adaptive averages
  6. 1995Building momentum-scaled adaptive moving averages
  7. 1995Adaptive length as a construction choice inside exponential smoothing
  8. 1998Testing price-channel breakouts with a lag-aware adaptive average
  9. 1998Constructing filters by nesting offsets and variable weights
  10. 1998Constructing an efficiency ratio adaptive average and entry filter
  11. 2001Encoding candle structure as a numeric filter
  12. 2001Adaptive averages driven by cycle-phase speed
  13. 2005Constructing an adaptive moving average from a fractal-dimension weight
  14. 2005Range-dimension adaptive exponential filter
  15. 2010Constructing simple, exponential, and adaptive averages
  16. 2010How a price-hugging smoother is assembled from ordinary averages
  17. 2013Evaluating an adaptive moving average against a same-window moving average
  18. 2016Three-layer confirmation: adaptive average, stochastic relative strength index, and stop-and-reverse
  19. 2017One-alpha reverse-path exponential smoothing
  20. 2018Pair two adaptive averages to filter swing turns
  21. 2018Two Adaptive moving averages as a confirmation pair
  22. 2018Constructing an adaptive filter for adoption-cycle reversals
  23. 2018Constructing a deviation-scaled adaptive moving average
  24. 2020Walk-forward and adaptive averages as two tests of the same trend
By fieldSystem design117 readingsBy year202085 readings
All 28 readings tagged Adaptive moving average
Also on Adaptive moving average5 readings
1991Building variable-length moving averages from partitioned price changes1991Variable-length moving average from change dispersion1992Constructing volatility-adaptive exponential smoothing1992Constructing volatility-adaptive trailing stops1995Building momentum-scaled adaptive moving averages