Passed its test, no edge

The drawdown brake: passed its test, no edge

· 8 min read

The idea

Hold each fund only while its price is above its 200-day average, and cash otherwise. Then scale the whole book down when recent volatility is high, so the portfolio aims for about 10% yearly volatility. We call it the drawdown brake. Moreira and Muir (Journal of Finance, 2017) showed that cutting risk when last month's volatility was high raised Sharpe ratios on several standard factors. The 200-day filter is a classic trend rule. Our 10% yearly volatility target is a fixed level we chose; their paper scales by the inverse of last month's variance.

What we did

  • Rule written first, never tuned. 200-day average, monthly (every 21 sessions) check, 10% volatility target from the prior month's realized volatility, no leverage.
  • Original test: 23 ETFs, two windows chosen in advance: 2017–2021 and 2022–2023. It was one of seven strategies in a single pre-registered batch (15 official trials).
  • Out-of-sample test: the same rule, unchanged, on seven groups of markets it was never tuned on: commodity ETFs, bond ETFs, real-estate ETFs, country ETFs, currency ETFs, four large crypto assets, and six major currency pairs. Pass rule written in advance: Sharpe at least that of equal-weight buy-and-hold, and worst drop at most 60% of buy-and-hold's. Confirmed if at least 4 of 7 groups pass.

What happened

On the original ETFs, against the plain 200-day rule it was written for, it edged ahead: Sharpe 1.00 against 0.97 in 2017–2021, worst drop 11.3% against 13.4%. In 2022–2023 it was about even (Sharpe −0.03 against −0.05). Across seven new market groups the same registered rule passed 6 of 7.

The fairer comparison is whether it beat simply holding the market on the days it was invested. It did not. In 2017–2021 its Sharpe was 1.00 against 1.04 for SPY held on the same days; in 2022–2023 it was −0.03 against 0.02. The smaller drops come from sitting in cash, not from picking better days. Fully invested SPY earned 18.28% a year in 2017–2021; the brake earned about 8%.

Across the seven new groups, the registered result vs fully invested buy-and-hold:

Market group Sharpe, brake vs hold Worst drop, brake vs hold Yearly return, brake vs hold Average invested
Commodities 0.28 vs 0.01 24% vs 68% 3.0% vs 0.0% 31%
Bonds 0.49 vs 0.42 10% vs 20% 2.8% vs 3.7% 48%
Real estate 0.36 vs 0.21 11% vs 51% 3.4% vs 3.3% 31%
Countries 0.29 vs 0.29 (missed by 0.008) 23% vs 61% 3.6% vs 5.6% 48%
Currency ETFs 0.14 vs −0.58 5% vs 13% 1.5% vs 0.0% 28%
Crypto 1.55 vs 1.24 23% vs 88% 28% vs 101% 17%
Currency pairs −0.12 vs −0.22 15% vs 34% 0.8% vs −0.5% 49%

Compared with holding less

The numbers below are the same-days and bumpiness-matched controls from the research report. They are not a new official pass/fail. The same-days and constant-weight mixes are frictionless, which slightly favours the control.

On the original 23 funds, against SPY:

Window Brake (Sharpe / yearly return / worst drop) SPY on the same days SPY at the same bumpiness
2017–2021 1.00 / 8.05% / 11.3% 1.04 / 10.0% / 15.3% 0.92 / 7.5% / 13.2%
2022–2023 −0.03 / 3.41% / 6.1% 0.02 / 3.67% / 6.7% 0.12 / 4.31% / 6.2%

Fully invested SPY was 18.28% a year in 2017–2021 (Sharpe 0.92) and 4.26% in 2022–2023 (Sharpe 0.12). Against SPY held on the same days, the brake did not add Sharpe or return in either window. Against a constant-weight mix with the same bumpiness, it had a higher Sharpe in 2017–2021 (1.00 vs 0.92) and a slightly higher yearly return (8.05% vs 7.5%); it lagged in 2022–2023.

Across the seven new groups, Sharpe versus the same-days basket (again, not a new pass/fail):

Market group Brake Same-days basket Fully invested basket
Commodities 0.28 0.29 0.01
Bonds 0.49 0.69 0.42
Real estate 0.36 0.33 0.21
Countries 0.29 0.29 0.29
Currency ETFs 0.14 −0.44 −0.58
Crypto 1.55 1.54 1.24
Currency pairs −0.12 −0.03 −0.22

On bonds, most of the Sharpe gain versus a fully invested basket disappears once you only hold the basket on the same days. On currency ETFs the brake still looks clearly better. A constant scale does not change Sharpe when idle cash earns the T-bill rate, so the bumpiness-matched Sharpe equals the fully invested Sharpe; we do not repeat it as a third column.

The official runs

These are the official default-cost results from the pre-registered literature batch. Sharpe and yearly return credit idle cash at the 3-month T-bill rate, the same way as the figures already on this page. Worst drop is from the simulated book as traded. The same-days and fully invested SPY rows are reference points, not trials in the batch; the SPY mixes are frictionless, which slightly favours the reference.

The 2005–2016 window was used to form the 200-day average. This rule was not an official pass/fail run on that window, so it is not in the table.

2017–2021SharpeYearly returnWorst dropTime invested
Drawdown brake Pre-registered · Confirmed1.008.05%11.3%57%
Plain 200-day overlay0.978.97%13.4%65%
SPY, same invested days Reference1.0410.0%15.3%57%
SPY, fully invested Reference0.9218.28%33.7%100%
2022–2023SharpeYearly returnWorst dropTime invested
Drawdown brake Pre-registered · Confirmed−0.033.41%6.1%39%
Plain 200-day overlay−0.053.28%6.1%39%
SPY, same invested days Reference0.023.67%6.7%39%
SPY, fully invested Reference0.124.26%22.1%100%
Growth of one dollar for the drawdown brake, SPY held on the same days, and fully invested SPY, in 2017–2021 and 2022–2023.
Growth of $1. The brake credits idle cash at the 3-month T-bill rate, the same way as the yearly returns above. The SPY mixes are frictionless. Official windows only; the January 2022 embargo between them is not shown.
Drawdown from peak for the drawdown brake, SPY held on the same days, and fully invested SPY, in 2017–2021 and 2022–2023.
Drawdown from peak. The brake uses the simulated book as traded. The smaller drops come with a book that is often in cash, not from beating SPY on the days it was invested.

The brake was one of seven ideas, and 15 official default-cost trials, written down before this batch ran. It was the only idea that passed its own pre-registered rule on both confirmatory windows. It still failed the batch-level deflated-Sharpe check (0.78 in 2017–2021 and 0.40 in 2022–2023, against a 0.95 bar). After it confirmed, we also ran it at zero costs and at double the usual spread: 2017–2021 Sharpe stayed 1.01 and 1.00; 2022–2023 stayed −0.02 and −0.04. Those cost checks were not part of the official 15 trials.

2017–2021, official batchSharpeYearly returnWorst dropTime investedOwn rule
Turn of the month0.753.82%6.7%12%Fail
Pre-Fed drift−0.300.70%5.1%1.9%Fail
Turn of the month plus Fed days0.713.74%6.7%13%Fail
Drawdown brake Confirmed · Highest Sharpe1.008.05%11.3%57%Pass
Vol-targeted core funds0.8911.54%18.3%77%Fail
Yield-curve gate0.998.78%10.0%58%Pass
Moon phases0.424.52%16.6%38%Sign test only
2022–2023, official batchSharpeYearly returnWorst dropTime investedOwn rule
Turn of the month−0.322.87%4.6%7.0%Fail
Pre-Fed drift0.314.16%1.8%1.1%Pass
Turn of the month plus Fed days−0.342.74%5.2%7.7%Fail
Drawdown brake Confirmed−0.033.41%6.1%39%Pass
Vol-targeted core funds Highest Sharpe0.356.87%10.7%55%Pass
Yield-curve gate−0.671.88%4.2%7.5%Fail
Moon phases0.175.03%10.8%38%Sign test only

The yield-curve gate passed 2017–2021 and failed 2022–2023, so it was not confirmed. The vol-targeted core book passed 2022–2023 (and had the highest Sharpe in that short window) but failed 2017–2021, so it was not confirmed either. Moon phases were judged on a close-to-close sign test, not on the executable Sharpe shown here. Each of the other ideas has its own write-up in the graveyard.

Read this part carefully

  • Sitting out, not better days. In the strong 2017–2021 stock market it earned about 8% a year against 18.28% for fully invested SPY, and it did not beat SPY held on the same days. That is the cost of a book that is often in cash, not evidence that the rule picked better days. In crypto it earned 28% a year against 101% for buy-and-hold, because it was invested only 17% of the time.
  • It still fails the batch-level overfitting check. The deflated Sharpe — a probability, not a marked-down Sharpe ratio, that asks how believable the best result still is after the ideas in the batch — was 0.78 for 2017–2021 and 0.40 for 2022–2023, against a 0.95 bar. The idea was confirmed under its own pre-registration rule, not by clearing that check.
  • Countries missed. The Sharpe there was 0.29 against 0.29 for buy-and-hold, short by 0.008. We did not move the line to make it 7 of 7.
  • Short, friendly samples. Crypto history here is roughly 2016 to 2023, a period of enormous growth. The 2022–2023 stock window is about 480 sessions, too short to say much.
  • Paper only. Simulated fills and costs. No real money has used it. Prices come from free data and the fund list is not survivorship-free.
  • Not advice. Do not read this page as a reason to buy anything.

Receipts: the rule and pass criteria were written down and committed before the batch ran, and again before the cross-market test. The data for 2024 onward was not used.