Frontier Model: An Open, Walk-Forward Validated Bitcoin Trading Model (XGBoost, Full Track Record Published)
I want to share a project I've been building: a systematic Bitcoin trading model where every trade, the full methodology, and the failed experiments are published in the open — not just the wins.
What it is
Frontier Model is an XGBoost strategy trading BTCUSD on daily candles, using 8 inputs derived from price structure, realized volatility, and normalized momentum/range behavior. No discretionary overrides. Notably, it's not an LSTM or any recurrent architecture — those were tested extensively earlier in the research and didn't survive the same validation bar the current model had to clear.
How it was selected (this is the part most projects skip)
Frontier Model was selected using VAL + VAL2 scores only — never TEST — out of 17,999 completed search trials. Once selected, it went through separate walk-forward and Monte Carlo confirmation:
V83 stitched walk-forward: positive in 8/8 windows, beat the paired baseline in 8/8
V84 final confirmation: positive in 6/6 windows, beat baseline in 5/6, with 91/128 paired seeds winning and the Monte Carlo gate passing 128/128 seeds
TEST was only looked at after all of that was locked in — the way it's supposed to work, and rarely how it's actually done.
The numbers — TEST vs TRAIN, side by side, on purpose
Full research track: January 2018 – September 2026, 733 trades, +759 simple percentage points (~+87.3%/yr simple annualized — a research statistic, not a compounded account return).
TEST (out-of-sample) TRAIN (in-sample)
Period Feb 2025 – Sep 2026 Jan 2018 – Jul 2022
Trades 174 294
Annualized return (simple) +66.1%/yr +106.0%/yr
Win rate 26.4% 37.1%
Max drawdown 4.8 pt 8.4 pt
SQN 3.26 —
Win rate is low — 31.1% across the full track, 26.4% in TEST. That's by design: many small losses get offset by a smaller number of much larger take-profit wins. If low win rate makes a system look broken to you, that's worth sitting with, because it's how a lot of legitimately edge-positive systems actually look.
Backtest includes 0.045% commission per side. Swap, funding, and live execution slippage are not included — so a real or simulated account can differ from this research curve, and I say that on the site itself, not just here.
What's independently verified vs. what's a backtest
The numbers above are a research/model study. Separately, there's a live-tracked simulated account, read directly by FXBlue (a third party with no reason to flatter it) — the two are shown together but are not the same dataset, and shouldn't be read as if one proves the other.
Everything is open
New signals posted to Telegram at execution, before the outcome is known
Full methodology, feature construction, walk-forward writeup, and the rejected experiments (including why LSTM lost): on the site
Research archive and code: GitHub
Technical questions (validation, feature design, why something was dropped): Discord
No paid tier, no private signal channel, no broker affiliate, no copy-trading service.
Standard disclaimer, because it needs to be said plainly: this is not investment advice. Bitcoin trading carries high risk and you can lose some or all of your capital. Backtested and simulated results have real limitations and do not reliably predict future performance. Nothing here is a recommendation — decide for yourself, and don't risk money you can't afford to lose.
Rigorous approach to walk-forward validation and avoiding data leakage by keeping the test set isolated until the end. Dropping recurrent models like LSTM based on empirical performance rather than hype is also a solid engineering choice.
Thanks, really appreciate that — and yeah, the LSTM thing stung a bit at the time, ngl. Sunk a good chunk of weeks into architectures I was pretty sure would win just because "deep learning" felt like the more serious answer. Turned out the extra complexity was just extra ways to overfit on daily candles that don't have that much signal packed into them to begin with, plus training runs that weren't even reproducible run to run — hard to trust a model you can't get the same number out of twice.