strategy deep-dive: lyra-gold
alpha micro-digest
lyra-gold: the sizing fix that saved a strategy
lyra-gold started as a mean-reversion bot on gold futures. the idea was simple. gold overextends intraday, snaps back, and you collect the spread. entry on a z-score threshold, exit on reversion to a rolling mean. one minute candles. one contract. nothing exotic.
the first backtest looked promising. profit factor above two. max drawdown under three percent. but when i ran the phase D audit, the fold test tore it apart. train on one month, test on the next, repeat across the full window. the profit factor held in some folds, collapsed in others. fold rate was sixty percent. not good enough.
the problem was position sizing. the original version used ATR based stops and a fixed risk per trade. gold volatility shifts regime to regime. a stop calibrated for a quiet month gets shredded in a volatile one. the bot was betting too big when gold was jumpy and too small when it was calm. the signal was fine. the sizing was wrong.
so i ripped out the ATR logic. replaced it with equity relative sizing. each trade risks a fixed fraction of current equity, not a fixed dollar amount. the stop distance still adapts to volatility, but the dollar risk scales with account size. in backtest, the fold rate jumped to ninety two percent. profit factor settled at two point one five. max drawdown stayed under two percent. the strategy stopped being a volatility regime bet and became a pure mean-reversion play.
live results since the fix. two hundred seventy seven trades. win rate twenty five point three percent. profit factor two point four two. total pnl five thousand two hundred twenty five dollars on fourteen thousand seven hundred ninety five dollars equity. max drawdown two point zero five percent. the win rate is low, which is normal for a reversion strategy with wide stops and tight targets. the profit factor is what matters.
the bot runs fully automated. i do not touch the parameters. it trades gold futures on a one minute timeframe. entry when price deviates beyond a z-score threshold from a rolling mean. exit on reversion or a time stop, whichever comes first. the time stop prevents capital from getting stuck in a trend day. gold trends sometimes. the bot is not designed for that.
what could be overfit
the z-score threshold and the rolling mean window were optimized on a specific backtest period. gold’s microstructure changes. if the exchange adjusts tick sizes or trading hours, the mean-reversion pattern could weaken. the fold test helps, but it is still in sample optimization across folds, not true out of sample. the equity relative sizing fixed the regime sensitivity, but the entry signal itself might degrade if gold’s intraday behavior shifts structurally.
what would falsify it next
first, a live drawdown exceeding four percent. the backtest max drawdown was two point zero five percent. if live drawdown doubles that, the strategy is likely broken. i would halt it immediately and investigate.
second, a rolling thirty day profit factor below one point two. the backtest profit factor was two point one five. some degradation is expected live. but if a full month of trading produces a profit factor under one point two, the edge is probably gone. i would pull the bot and re audit on recent data.
lyra-gold is not a home run. it is a small, boring edge that compounds. the sizing fix turned a fragile idea into something that survives regime shifts. for now, it works.
more at falsifylab.com
— research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.