Why a Quiet Strategy Is Worth Chasing

Most backtests brag about returns. This one, shared in December 2013 on the Systematic Investor blog and credited to a reader named Pierre Chretien, brags about something less glamorous: turnover. The goal was a strategy that trades rarely, holds up across several look-back windows, and still makes money.

That goal matters more than it sounds. Every trade costs a spread, a commission, sometimes a tax bill. A portfolio that rebalances its way through 150% annual turnover can lose a full percentage point or more to friction before any skill shows up, and a backtest that ignores this flatters itself.

The setup used ten funds spanning US equities (VTI and its older mutual fund history), Europe, Japan, emerging markets, real estate, commodities, gold, and three maturities of Treasuries. Data began in 1998, with monthly rebalancing and a cap of 60% in any one product. Four risk-based allocation methods were tested: equal weight, minimum variance, minimum correlation, and random subspace versions of the latter two.

The question was simple. If we smooth the inputs that feed the optimizer, and smooth the momentum signal that decides what is eligible, does the portfolio get calmer without getting worse?

Two Kinds of Averaging, One Shared Worry

Every quantitative strategy has to pick a look-back window. Do we use 40 days of returns to estimate risk, or 60? Do we use 120 days of momentum, or 250? Usually we pick whatever worked best in the sample, which is a polite way of describing overfitting.

Averaging sidesteps the choice. Instead of betting on one window, we blend several and let no single one dominate.

The first technique applies to input assumptions, meaning the expected returns, volatilities and correlations that go into the optimizer. In the code, we stack the returns from 20, 40 and 60 day windows together, with a one day lag, and the recent observations naturally appear in more of the windows. That gives recent data more weight without discarding the older stretch.

The second technique applies to momentum. For each look-back (20, 60, 120 and 250 days, with a three day lag), the code computes the price ratio, scales it to an annual rate by multiplying by 252 divided by the window length, and averages the results. A fund is eligible only when that averaged figure is above zero.

This is absolute momentum, not the cross-sectional kind. We judge each asset against its own past, as in Gary Antonacci's work and the Flexible Asset Allocation paper by Keller and Van Putten. We are not ranking funds against each other; we are asking whether each one is rising or falling.

One detail deserves attention: the annualizing step. A 20 day gain of 2% is a much stronger signal than a 250 day gain of 2%, and when we scale by window length we put them on the same footing before we take the average.

Six Strategies and Four Questions

The experiment crosses two choices. Input assumptions can be plain (a single 252 day window) or averaged. Momentum can be absent, a single 120 day filter, or the averaged version. That produces six variants, each run through the five allocation algorithms.

The author then compared them in pairs, and each pair answers a separate question. Does a momentum filter help at all? Does averaging the inputs reduce turnover? Does averaged momentum beat the plain 120 day filter? And does the fully averaged combination beat the simple momentum version?

The comparison is a good template because it isolates one change at a time. When we stack five tweaks into a final strategy and only show the end result, we can never tell which tweak earned its place. Pairwise testing exposes the ones that did nothing.

A caution on method: the universe here has ten assets and about fifteen years of history, including the 2000 to 2002 bear market, the 2008 crash, and a long bond rally. Those episodes dominate the results. A filter that happened to step aside from equities in 2008 will look brilliant, and we should be skeptical of how much of that is repeatable.

What the Comparisons Showed

The first result was the clearest. Adding a momentum filter improved every allocation algorithm tested. Equal weight, minimum variance, minimum correlation and both random subspace versions all did better with the filter than without it.

That makes intuitive sense. Risk-based optimizers know nothing about direction. Minimum variance will happily load up on an asset that is quiet because it is slowly collapsing. A trend filter removes those holdings before the optimizer sees them.

The second result concerned turnover. Averaging the input assumptions reduced it, which was the stated purpose. Smoothing the covariance estimates means the weights shift less from month to month, so fewer trades are needed.

The third was murkier. Swapping the single 120 day momentum for averaged momentum produced mixed results across the algorithms. Some improved, some slipped. There was no clean winner, and the honest reading is that the averaged signal is about as good as the plain one, with the benefit being robustness to the choice of window rather than higher returns.

The fourth comparison pitted plain input assumptions with simple momentum against the fully averaged version. The source text cuts off before the conclusion, so we will not claim a result. What we can say is that the logic of the design points toward similar returns with lower trading.

Where This Leaves a Real Portfolio

The test universe is made of broad funds, which is how most investors can implement it in practice. VTI stands in for US stocks, and the same approach works with any liquid ETF. A factor fund like QUAL could replace the broad market holding, and the filter would treat it the same way: in when its averaged momentum is positive, out when it turns negative.

Individual stocks are a different matter. Take a bank like MS or C during 2008. A 250 day momentum window would have flagged weakness well before the worst of the decline, while a 20 day window would have flipped back and forth through the rebounds. Averaging the windows would have softened both problems, though no look-back mix avoids the pain of a gap down on a single bad day.

Now consider a holding like META, which fell sharply in 2022 and then recovered more than most expected. A filter built on 250 day returns would have exited late and re-entered late. For a single name, the cost of that lag can be large. The strategy in the source is built for diversified funds, where a late exit from one asset is offset by what else the portfolio holds.

Here is a rough example. Suppose a 60% cap on any one product means the biggest allowed position is 60% of the portfolio. If US equities turn negative on the averaged signal, that sleeve drops out and the optimizer redistributes the money across whatever else is eligible, perhaps Treasuries and gold. The cap prevents the optimizer from dumping everything into a single survivor.

The main risk is whipsaw. A market that falls 8% and then recovers within a month will trigger an exit and a re-entry at a worse price. Averaging reduces how often this happens but cannot remove it. The main opportunity is the opposite case: a slow, grinding decline of the 2000 to 2002 type, where a filter protects capital over many months.

Putting the Idea to Work

Anyone who wants to try this should start with the data. The author spliced mutual fund histories onto ETF histories (VTI with VTSMX, for example) to push the sample back to 1998. That is a sensible trick, but the two series are not identical, and fees and tracking differences can nudge results.

Timing is straightforward: monthly. The signal is computed at month end and trades happen at the next opportunity. The lag of one day on the input assumptions and three days on the momentum is not decoration. It prevents the strategy from using information it would not have had in time to act on.

For readers who do not code in R, the same logic fits in a spreadsheet. Compute the price ratio for each fund over 20, 60, 120 and 250 days, multiply each by 252 over the window length, average the four, and hold only funds where the result is positive. Then weight the survivors equally, which is the simplest of the five algorithms tested and still benefited from the filter.

A few practical challenges come up. Taxable accounts will feel turnover more than the backtest suggests, so the lower trading from averaging is worth more there. Funds with redemption fees or wide spreads can erase the gains from a monthly rule. And a strategy that sits in cash or short Treasuries for long stretches will lag a rising stock market, which tests patience more than any statistic.

What to Take Away and What to Do Next

Averaging is a way of admitting we do not know the right window. That admission costs little and protects against the most common failure in systematic investing, which is a rule that worked on the past because we fitted it to the past.

The evidence in this experiment supports two modest claims. A momentum filter helped every allocation method we tested, and averaged input assumptions cut turnover. The advantage of averaged momentum over a single well chosen window is weaker, and our own results were mixed.

For a reader who wants to act on this, a sensible sequence looks like this. First, we build the averaged momentum signal for a handful of broad ETFs and check what it would have said at the end of 2007 and 2008. Second, we compare turnover and drawdown against a plain buy and hold version of the same portfolio. Third, we change the windows slightly (say 30, 90, 150 and 240 days) and see whether the results hold. If they fall apart, we fitted the strategy to the original numbers.

Finally, size the strategy to the account. A 60% cap per product, as in the source, is a decent starting guardrail. Keep the universe diversified enough that one wrong signal costs a few percent rather than a quarter of the portfolio, and treat the backtest as a reason to run the idea, not a promise of what it will deliver.