Textbook Trend-Following on the Wrong Instrument: Why Donchian Channels Failed on Gold¶
Why the board suggested Donchian¶
After our first attempt to fill the US-session gap failed, the board ran a ChatGPT research sweep to get independent strategy recommendations. Three candidates came back: ADX-Gated Session Breakout, NY Open Intraday Momentum, and Donchian Channel Breakout.
Of the three, Donchian looked the most promising. Trend-following on commodities is a well-studied archetype, and Gold is the textbook example of a market that trends persistently enough to reward channel breakouts. The Donchian Channel itself is elegant: it simply trades the breakout of the highest high and lowest low over a lookback period, assuming that any genuine trend will push into new N-bar territory.
It seemed obvious. Gold was supposed to be the answer.
The hypothesis: harvest trends on the right instrument¶
The intuition was straightforward. Unlike the equity index and FX pairs we'd tested before, Gold (XAUUSD) is designed to trend. It's a safe-haven asset that responds to macro shocks, geopolitical events, and Fed policy shifts with sustained multi-bar moves. A Donchian channel breakout—trading the first breakout of an N-bar range—should capture the opening of those moves.
We tested two channel periods: a faster short-lookback variant (designed for intraday moves) and a slower long-lookback variant (designed to filter out smaller chop). Both were paired with a trending regime filter, volatility-based stop/take-profit sizing, and standard exit logic.
Testing: margin above threshold, but negative¶
The discovery period covered several years of intraday data with spread-aware fills and a standard performance threshold to promote candidates to out-of-sample testing.
Both channel periods failed to meet it. The faster short-lookback variant was a high-frequency loss machine—thousands of round trips, unprofitable on every metric. The slower long-lookback variant was less bad but still failed across the board.
The gap between the two periods was real, but it moved in the wrong direction. Slowing the channel down made the strategy less bad, not good. There was no nearby parameter adjustment that would have fixed it.
Why it failed: microstructure doesn't trend¶
This was harder to understand than the first failure. Gold does trend. The market is designed for this archetype. But the intraday microstructure at the resolution we needed to fill the US-session gap doesn't exhibit the kind of sustained directional moves that Donchian breakouts require.
What we were seeing was exactly the kind of behavior that kills breakout systems: many small reversions, fewer large trends, and a high proportion of false starts that never develop into the full N-bar move that could justify holding the position.
The issue wasn't Gold itself. It was Gold at the intraday resolution we were trading, during the specific session hours. Longer timeframes might have shown different behavior. Different session windows might have shown different behavior. But in the exact market-time-resolution combination we needed, the Donchian archetype didn't have edge.
What changed after¶
We added another layer to our research process: validate not just the strategy archetype, but the market-timeframe-session combination it's supposed to work on.
Gold is a trend-following market. Donchian channels are a trend-following strategy. But the conjunction of Gold + intraday resolution + US session hours didn't produce the Donchian edge in the data. We could have spent weeks tuning parameters, adding filters, or trying different exit logic. Instead, we documented the microstructure mismatch and moved on.
Later, when every Phase 4 and Phase 5 candidate we tested on the US session hit the same wall, we started to see a pattern. The gap wasn't about the specific strategy. It was about the microstructure of the market during those hours.
Radius Red learns in public. This post is part of a series on research that didn't make the portfolio—a window into how systematic trading research actually works, failures and all.
Series: The US-session gap is structural, not a tuning problem