Sector timing

An earlier test had already killed stock-picking in these sectors: the return lived in owning the sectors, not in choosing names. A first, unfitted cut of a 200-day trend rule then looked promising: a Sharpe ratio of 0.84 against buy-and-hold's 0.74, and the worst drawdown cut from −52% to −16%.

Timing results like that are easy to fool yourself with. Being in the market during an up-decade helps regardless of skill, so beating buy-and-hold is not enough. A real timing edge also has to beat random timing with the same amount of time in the market, and most timing edges die on that control.

Two equal-weighted baskets, energy and construction, built from 119 stocks, with daily returns clipped at ±50% to guard against corrupt prices. Cash earns a T-bill proxy (the BIL ETF), not zero, so being out of the market is neither unfairly punished nor rewarded in high-rate years. The walk-forward produced 108 out-of-sample months. The ETFs XLE, XLB and PAVE were named in the plan as an independent cross-check.

Everything was frozen before the first run. Monthly rebalancing, 10 basis points per side on every change in exposure, and a deliberately small grid of about 42 configurations: moving-average, momentum and breakout rules, with or without a volatility target, and three ways to split money between the two sectors.

Anchored walk-forward, refit yearly on at least five years of training: in each fold, pick the top 20% of configurations by in-sample Sortino ratio, average their positions, and trade that ensemble on the unseen year. The verdict is the single concatenated out-of-sample curve, never the configuration that happened to win.

The edge counted as real only if it (a) beat buy-and-hold's Sharpe by a clear margin, (b) beat 5,000 random in/out schedules with the same time in the market at p < 0.05, using a block bootstrap, (c) held in at least 2 of 3 subperiods, not just the 2020 crash, and (d) replicated on the tradable ETFs.

Out of sample it looked good at first: a Sharpe ratio of 0.93 against buy-and-hold's 0.86, and half the drawdown, −24% against −52%. It also made less money: 16.3% a year against buy-and-hold's 23.1%.

Then the decisive control. Keeping the strategy's exact positions but shuffling which dates they landed on, 5,000 times, 26% of random schedules did as well (p = 0.26). The confidence interval on the Sharpe advantage, [−0.22, +0.42], straddles zero, and timing beat buy-and-hold in only 1 of 3 subperiods. The same benefit comes from a fixed mix of 65% stocks and 35% cash, with no trend rule at all.

The 2025–26 lockbox has been spent, so the only clean test left is forward paper-trading. The sample is also trend- and crash-heavy, which flatters trend-following; any new idea should be checked in a choppy market first. A later look at short-horizon momentum did exactly that, and in the neutral 2014–2019 market its apparent edge largely disappeared.