Joshua Dunnink · Trading Systems

← Dunlin Gold

Dunlin Gold - research note (every number names its ARM, FEED and WINDOW)

All figures: MetaTrader 5 Strategy Tester, real ticks, $1,000 deposit, default lot (0.01 per $1,000 of balance), 150 ms delay, leverage 1:500, version 0.23 defaults (no .set), window 2025-09-01..2026-11-14. Arm: "Defaults", version 0.23. The IC Markets and VT Markets rows replay the same version's trade lists at a fixed 0.01 lot from $1,000.

What it is

Gold neural-network window trader. A small neural network reads the last two and a half days of 15-minute candles and decides whether gold is in an orderly, range-like stretch with a lean up or down; if so it opens a trading window and takes up to three positions on that side. Every position has a stop loss from the moment it opens. Profit is protected in stages: the stop moves to entry after a small favourable move, then locks profit and trails. No martingale, no grid, no averaging down.

How the network was made

It was trained on 2024-12 to 2025-06-30 gold and on many copies of that market: the same days with the clock shifted, older gold history (2013 onward) rescaled to today's price, time-stretched and amplitude-matched, and about forty synthetic paths built from real days, each flipped upside-down with probability one half. Training stops at 2025-06-30; everything after is data the network never saw.

Result by feed

Feed Result in dollars Profit factor Largest dip Trades Result without the five percent biggest winners
BlackBull ← claim 1,027 2.02 8.05 % 712 725
IC Markets 1,496 2.66 5.58 % 710 989
VT Markets 1,136 2.24 7.75 % 724 857

The claim row is BlackBull, the weakest of the three on profit factor. The dip column is the closed-trade balance; on BlackBull the equity drawdown including open positions was 13.55 % (backtest, defaults, BlackBull, 2025-09-01..2026-11-14).

What the curve looks like

Reading it honestly