Horizon · Strategy dossier

Horizon Donchian

A breakout system that rests twelve stop orders just inside the edges of four Donchian channels, sizes every distance as a percentage of yesterday's close, and then manages every fill identically — so most trades exit on a trailing stop rather than at the target they were given.

12 strategies 4 shared channels XAUUSD · H1 Stop orders, never market Exits flat in S

Everything is measured in S

One number sets the scale of the entire system. Nothing is expressed in dollars or pips.

S = 0.5% × previous broker-day close

On gold at $4,400 that makes S = $22. Level offsets, stop distances, targets, break-even triggers and trailing gaps are all multiples of S. The consequence is that the EA rescaled itself automatically as gold went from $1,800 to $4,400 without a single setting being touched.

S at $1,800
$9.00
2022 regime
S at $2,600
$13.00
2024 regime
S at $4,400
$22.00
2026 regime
Known weakness

S tracks price, not volatility. In a quiet year the same multiples are far too wide, which is the leading explanation for why 2021–22 were flat-to-negative while 2024–26 were strong. Gold's range measured in S roughly doubled across those two halves.

How it finds levels

A Donchian channel is just the highest high and lowest low of the last N hours, taken on completed H1 bars. The EA runs four of them at once.

chanHigh = highest high of the last N hours chanLow = lowest low of the last N hours buy stop @ chanHigh + kBuy·S kBuy is NEGATIVE → just inside the top sell stop @ chanLow + kSell·S kSell is POSITIVE → just inside the bottom

It does not wait for the breakout — it front-runs it. With kBuy = −0.149 the buy stop sits 0.149·S below the channel high. On gold that is roughly $3 early. You are filled just before the channel actually breaks, which is the single most distinctive thing about the entry model.

Group A — the 1350-hour channel
Real XAUUSD H1 bars. The channel edge and the resting order level are separate lines — note the gap.
channel high / low buy stop sell stop

Four channels, four speeds

The twelve strategies do not each get their own channel. They share four, and the difference between a "slow" and a "fast" strategy is which of these it reads.

GroupLookbackIn daysStrategiesCharacter
A1350 h56.20, 3, 5, 8Eight-week channel. Rarely moves; a break is a genuine multi-month high.
B162 h6.81, 2, 4, 6, 7One-week channel. The bulk of the book.
C89 h3.79Half-week.
D17 h0.710, 11Under a day. Fires constantly, tiny edge per trade.
All four channels on the same bars
The 17-hour channel hugs price; the 1350-hour one barely moves. Same data, four different definitions of "the edge".
A · 1350h B · 162h C · 89h D · 17h

The twelve strategies

Each has a fixed stop and target in S, and a fixed lot weight. The weights and distances were fitted from 66,921 of the vendor EA's own orders — they are calibration constants, not preferences.

#GroupChannelkBuy SL (S)TP (S)R:RLot wtProfile
Why the win rate is high but the profit factor is not

The mix is deliberate. Strategy 3 risks 1.00 S to make 4.09 S. Strategy 7 risks 4.58 S to make 0.67 S — it wins around 90% of the time and loses badly when it loses. Blending runners and scalpers on the same channels is what produces an ~87% win rate at a profit factor of only 1.4–2.0. It is twelve bets on four channels, not twelve independent systems.

What happens after a fill

The stop and target you see on the order are mostly theatre. A flat trailing rule takes over almost immediately and does the actual work.

  1. The order restsA stop order sits at the level. Nothing happens until price rises (or falls) to reach it. The EA never enters at market.
  2. Fill, with SL and TP attachedStop at level − SL_S·S, target at level + TP_S·S, both fixed per strategy.
  3. Break-even at +0.35 SAs soon as the trade is 0.35 S in profit, the stop jumps to entry + 0.10 S — a small profit is locked. This distance is the same for all twelve strategies, not scaled to each one's target.
  4. Then it trails 0.80 S behindAlso flat across the book. This is what actually closes most trades.
  5. The ratchet, if it goes wrongIf the trade runs 0.45·SL against you, the target is pulled in toward price — settle for a smaller win if it comes back. Used on about 7% of positions.
A real fill, with the trail drawn
Strategy 0 on the 1350h channel.
entry original stop target trailing stop

Measured, not assumed

These rules were reconstructed from 3,561 "position modified" events in the vendor EA's own journal, then confirmed across 15,044 trades. The exit mix that produces:

Exit on original stop
1.6%
the stated SL almost never fires
Exit on original target
16%
strategy 3 hits its own TP 0% of the time
Exit on a moved stop
70.6%
in profit — the trail is the strategy

Housekeeping that matters

Daily wipe at 23:45

Every pending order is cancelled at 23:45 broker time and re-armed 80 minutes later, at 01:05. Copied from the vendor's ScheduledClose. It prevents stale levels triggering on the illiquid rollover.

Level freeze while filled

While a position is open on one side, that side's level is re-armed at the same price rather than recomputed. Measured from the vendor: 56% of re-arms after a fill are at an identical price, median 4.5 minutes later.

Never drag a level onto price

A rolling channel can drop a level onto the market and fill instantly — observed live at 15:00:00.247. A breakout must trigger because price rose to the level, never because the level fell to price.

Fixed lots, and a silent floor

Lots are per-strategy weights × one global multiplier, and do not grow with the account. Any strategy whose lot rounds below the broker minimum is silently dropped — which is why small accounts quietly stop running half the book.

The trap this creates on a small account

At a $2,600 balance with the shipped weights, strategies 1, 7, 8 and 10 fall under the lot floor and never trade — 8,141 of the vendor's 15,061 trades, 54% of the book. The account looks like it is running the strategy, and is running half of it.

Measured performance

Real ticks, XAUUSD, $10,000 deposit, one independent run per year. Two configurations: the full book, and the four-strategy Group A subset ("V8").

YearAll 12DDV8 (0,3,5,8)DD
What the years say

The full book returns roughly 3.2× what the four-strategy subset does, at roughly 3× the drawdown — and its losing years are much worse (−17.4% against −3.7%). Cross-strategy correlation of yearly P&L is 0.96, with 64 of 66 pairs above 0.90: the twelve are one bet expressed twelve ways. Selecting a "best" subset on early years and testing it on later ones underperforms simply running all twelve.