Signal research: a 147-rule sweep against buy and hold

Do optimized band and crossover parameters beat buy and hold out of sample? A parameter search always finds a winner in sample. The question is whether the winner keeps winning on later data, and whether its margin is larger than what the same search finds in data with no structure at all.

One large-cap stock, NVIDIA (NVDA): daily adjusted prices from Yahoo Finance, 1999-01-22 to 2026-08-07, 6,928 trading days after excluding the partial last day (recorded in the study). The prices are used under Yahoo's personal terms and are not redistributed; this page shows only statistics derived from them.

In sample 2000-01-21 to 2014-12-31 (3,760 trading days), out of sample 2015-01-02 to 2026-08-07 (2,916 trading days). Both halves contain a major drawdown and a major run.

147 rules in 6 indicator families: EMA crossovers (50), MACD (27), Bollinger bands (40) and Keltner channels (18), each traded both as a breakout and as a reversion, RSI (9), and on-balance volume against its own average (3). Sweeping both directions for the band families counts a choice researchers usually make after the fact.

Long or flat, 5 basis points per round trip, and Sharpe ratios with a zero risk-free rate. Each signal is computed at the end of one day and held through the next.

Rules are chosen on the in-sample half only, and the out-of-sample half is scored once. Two further checks test the search itself: 1,000 reruns of the whole 147-rule search on resampled returns, single days and 21-day blocks, to measure the best edge that searching alone produces; and a 22-year walk-forward that re-picks the best rule each year from prior data only and trades it the following year.

0 of 147 rules beat buy and hold on out-of-sample return. On Sharpe ratio, 13 of the 142 rules that traded out of sample did, all of them EMA crossovers, and none of the 13 ranked better than 21st of 144 in sample, so none would have been chosen. The in-sample pick of each of the 8 family and direction groups lost on both return and Sharpe (0 of 8 beat it).

The best out-of-sample Sharpe edge, +0.081 from EMA(8,200), is below the median best edge the same search found in noise: +0.119 resampling single days and +0.168 resampling 21-day blocks (recorded in the study).

Within EMA crossovers, in-sample and out-of-sample ranks ran in opposite directions (rank correlation -0.659): picking the in-sample best did worse than picking at random. Across all rules the two Sharpe ratios do correlate (+0.664), but that is exposure, not skill: out-of-sample time in the market tracks out-of-sample Sharpe at +0.892, and a rule's time in the market barely changes between the halves (+0.970).

The walk-forward, re-picking every year, compounded at 19.5% a year from 2005 to 2026 against 38.3% for buy and hold, and beat it in 5 of 22 years (recorded in the study).

One qualified positive, with its four caveats. The 8 EMA rules with a 200-day slow leg had out-of-sample maximum drawdowns between -43.2% and -55.6%, against -66.3% for buy and hold. But the best of them still lost on return (66.0% a year against 70.1%); none was selectable in sample (ranks 52 to 82); the best Sharpe margin among them (+0.081) is below the noise median; and it is one stock over one out-of-sample period. That is a trade of return for a smaller drawdown, not a signal edge.

28 correctness gates ran before the sweep (recorded in the study), including a mechanical no-lookahead proof: every input after a cutoff day was multiplied by random factors, and all 147 signal series were identical up to that day.

The in-sample best failed to beat buy and hold out of sample on both Sharpe and return at all 7 split dates tried, 2008 through 2020 (recorded in the study).

On this site, every number from the results table is recomputed from its 147 rows by a test in the test suite, so the page cannot drift from the data.

One stock and one out-of-sample period: a single draw, not a distribution. NVIDIA trended strongly upward over the out-of-sample window (buy and hold returned 70.1% a year), and a long-or-flat rule gives up return whenever it is out of the market, so the result should not be extended to a sideways or mean-reverting instrument.

Long or flat only: no short side, leverage, position sizing, or stops. Daily data only. 5 basis points is optimistic for the early years, when spreads were wider, which only strengthens the negative result. RSI was tested only in its usual mean-reversion direction, and the zero risk-free rate flatters every rule and buy and hold alike.