Outsample

Verdict series · Episode 01 · 2026-08-19

SuperTrend, as commonly traded.

BTCUSDT · 1D · 2019-01-01 to 2026-08-18 · params (10, 3.0), TradingView defaults, unoptimized · long/cash · costs 0.05% fee + 0.05% slippage per execution · bands v1 · seed 20260819

VERDICT: OVERFIT three independent kills, none of them close

SuperTrend's celebrated Bitcoin track record does not survive testing. Its equity path was one lucky ordering of trades. It lost money the one year that mattered. And a quarter of randomly shuffled histories, with all structure destroyed, beat it. The returns are bull-market drift, not timing skill. This is an evaluation of historical behavior, not a prediction, and that honesty is the product.

What we tested

The TradingView built-in SuperTrend (hl2 basis, Wilder ATR), default parameters, traded the way it is commonly traded: long when the trend flips up, cash otherwise, signal on the confirmed close, position from the next bar. Before any testing, the Python engine was proven against TradingView's own execution on identical bars: zero position disagreements across 2,787 bars, equity agreement to fifteen decimal places. Every number below runs on an engine that is exactly the public script.

The baseline already tells the story

7.83xstrategy multiple
17.05xbuy & hold multiple
0.58 vs 0.59Calmar, strategy vs holding
54%strategy max drawdown
Strategy equity versus buy and hold on a log scale, with losing periods shaded
Fig. 1 · Equity vs buy & hold, log scale. Shaded: the slices where the strategy lost money outright.

Kill one: the path was luck

Reorder the same 34 trades 400,000 times and read the drawdown distribution. 47.2% of orderings cut the account in half (the swing-class ruin depth). 99.4% breached the 30% abandonment depth, the point where a rational operator quits the system. The one historical path everyone backtests was among the gentlest available arrangements of its own trades.

Maximum drawdown distribution across 400,000 trade reorderings
Fig. 2 · Max drawdown across 400,000 trade-order resamples. The dashed lines are the abandonment and ruin depths.

Kill two: the bear year

One continuous run, sliced at year boundaries. 2022, the only sustained bear in the window, comes out at Calmar −0.98. The 2025–2026 slice is negative too. A trend strategy that only earns in bull years is a bull-market position with extra steps and a fee bill.

Kill three: noise beats it

Re-run the strategy on 15,000 shuffled histories: identical returns, volatility preserved, temporal structure destroyed. If SuperTrend's timing added information, the real result should sit far in the right tail. It sits at p = 0.239. About a quarter of structureless noise beat the real run. The seed is disclosed; shuffle it yourself.

Distribution of final multiples on shuffled histories versus the real result
Fig. 3 · Final multiples on shuffled histories vs the real result. Timing skill would live in the far right tail.

The twist: you cannot tune your way out

The parameter neighborhood is green everywhere: 100% of the ±20% grid is profitable, and the center is not a lucky peak (sharpness 0.87). This is the most useful finding in the episode: the settings were never the problem. Every parameterization captures bull drift, and none of them beats holding. The edge is absent, not mis-tuned.

Calmar ratio across the parameter neighborhood grid
Fig. 4 · Calmar across the parameter grid. Uniformly positive, uniformly beaten by doing nothing.

What this verdict does not say

Run it on your own chart

The episode ships with a free TradingView tool: SuperTrend Reality Check draws this exact comparison, cost-applied strategy equity vs buy & hold with no repainting and honest numbers, on any symbol you open. Publication pending on TradingView · @Outsample.

See the full method

Battery: baseline → 400,000-path trade bootstrap → year-slice walk-forward → ±10/20% parameter jitter → 15,000-run permutation, add-one smoothed p, seed 20260819. Verdicts are deterministic under bands v1. Research and education, not financial advice; no result shown implies future performance.

Intake

What are we doing for you?