The SMA Crossover, Honestly Backtested: QQQ and Bitcoin, Costs Included

Compute a fast and a slow simple moving average of the close. When the fast SMA sits above the slow one, the market is trending up — be long. When it drops below, be flat. No shorting, no leverage, no discretion. With fast = 50 and slow = 200 this is the golden cross of financial television fame, and a one-asset special case of time-series momentum — the effect Moskowitz, Ooi and Pedersen documented across 58 futures markets.

We run it on two deliberately different animals: the Nasdaq-100 ETF (exchange calendar, ~252 bars a year) and Bitcoin (trades every day of the week, ~365 bars a year). BTC has prices on weekends where QQQ has none, so each asset is backtested on its own calendar — per-asset dropna, per-asset annualisation — never forward-filled onto a shared grid.

The most important line in the backtest is not the signal. It is the shift. Today's SMA cross is computed on today's close— a price you cannot trade before you have seen it. So today's signal may only earn tomorrow's return. Drop the shift and the backtest buys every up-day one bar early: a lookahead bias that flatters almost any signal and quietly fabricates Sharpe.

def backtest(px, fast=50, slow=200, fee=0.001):
    signal   = (px.rolling(fast).mean() > px.rolling(slow).mean()).astype(float)
    position = signal.shift(1).fillna(0.0)   # <- no lookahead. The whole game.
    ret      = px.pct_change().fillna(0.0)
    cost     = position.diff().abs().fillna(0.0) * fee   # 10bp per side
    return (position * ret - cost).iloc[250:]   # SMA warm-up window

Two more discipline details. Every position change pays 10bp — entries and exits both. And the first 250 bars of each asset are reserved for SMA warm-up, so every parameter pair we test later is judged on the same evaluation window: 2015-12-30 → 2024-12-31 for QQQ, 2015-09-08 → 2024-12-31for BTC. The pandas core above is deliberately transparent; we rebuild the identical trades with vectorbt's Portfolio.from_signals as an independent referee — final value, max drawdown and trade count agree.

Growth of $100, log scale (BTC would render every other line invisible otherwise). Watch where the strategy detaches from buy-and-hold: always in the big drawdowns. A slow trend filter is not an accelerator — it is a brake pedal.

On QQQ the crossover ends at $439 versus $478 for buy-and-hold — it lost the return race — but it did so at 18.9% vol instead of 22.2% and cut the worst drawdown from -35.1% to -28.6%, for a slightly better Sharpe (0.97 vs 0.90). On BTC it kept pace with one of the great bull markets in modern data ($23,100 vs $38,954) while sitting out 36% of all days and trimming the max drawdown from -83.4% to -69.3%. Neither run is a money machine. Both are drawdown insurance — and the whole 10-year QQQ position changed hands just 5 times. Two conservatisms are baked into every number above and worth naming: flat days are credited nothing (the QQQ variant sits in cash 18%of the time — park that at the T-bill rate and the strategy line improves while buy-and-hold's cannot), and Sharpe is quoted on raw rather than excess returns. Both choices shade against the strategy, which is the right direction to be wrong in.

One backtest is an anecdote. Before believing 50/200, ask whether the neighbourhood agrees: we sweep fast ∈ {10, 20, 50, 100} × slow ∈ {100, 150, 200, 250} — 15 valid systems per asset, identical execution, identical costs — and heat-map the Sharpe. Both grids share one colour scale, so a glance tells you which asset rewarded trend.

This is the card's thesis in one picture. QQQ: every cell lands between 0.78 and 1.01, straddling buy-and-hold's 0.90 — only 7 of 15cells beat it, none decisively. Picking the best cell after the fact and calling it edge is selection bias, not alpha; on this asset the honest claim is "no return edge, a consistent drawdown brake". BTC: 12 of 15cells sit above buy-and-hold's 1.27, the whole grid stays in the 1.18–1.43 band, and every cell slashes the max drawdown. The same rule gets two different verdicts because trend behaves differently per asset— Bitcoin's decade delivered longer, cleaner trends and far deeper crashes for the filter to sidestep.

Costs are the second-biggest backtest killer after lookahead — but they bite in proportion to turnover. The 50/200 pair trades a handful of times a decade, so even 25bp per side barely dents it. Speed the system up to 10/100 and the toll booth opens:

Slow trend is nearly free to run. Fast trend pays a visible tax — QQQ's 10/100 variant loses 0.06 Sharpe going from free execution to 25bp, on 18 round-trip entries. Any strategy whose backtest only works at 0bp does not work.

Everything above is one ten-year window, evaluated in-sample. There is no walk-forward, no out-of-sample holdout, and no multiple-testing haircut for the 15 variants we just eyeballed per asset. 2015–2024 handed both assets two of the strongest trend decades they have ever printed — a regime gift the next decade owes nobody. Before promoting any cell of that grid, run the standard honesty checklist: point-in-time data, walk-forward splits, pessimistic costs, and a deflated Sharpe— Bailey & López de Prado's correction for exactly the selection bias a 15-cell grid search manufactures — on whatever looked best.