A market-neutral crypto strategy,
built by evidence, trading live.
Axon Ridge runs a machine-learned long/short book on Hyperliquid perpetual futures. It doesn't bet on crypto going up. It bets on which coins will do better than the others — and holds equal amounts long and short so the market's direction cancels out. Everything below is real: the research, the paper trail, and the live account.
We don't predict the market.
We rank it.
Every day, a machine-learned model looks at 50 of the most liquid perpetual futures on Hyperliquid and answers one question: over the next 24 hours, which of these coins will do best relative to the rest — and which will do worst? We buy the top of that ranking, sell short the bottom, in equal dollar amounts. That is the whole strategy. No leverage stacking, no directional bets, no discretionary overrides.
Imagine a horse race that runs every single day with the same 50 horses. You don't need to know whether it will rain — you only need to spot which horses are in form and which are limping. We back the 8 best-looking, lay the 8 worst-looking, in equal amounts. Whether the whole race runs fast or slow doesn't matter — only the gaps between horses do.
Evidence before capital
No strategy touches real money until it has passed a pre-registered three-stage research funnel, then traded on paper. The pass/fail rules are written down before the tests run, so we can't move the goalposts when results come in.
Market-neutral by construction
Longs and shorts always balance to the dollar. If Bitcoin drops 20% overnight, our shorts profit roughly what our longs lose. Returns come from relative mispricings, not from riding the market.
Costs are a first-class citizen
Every backtest number is net of realistic trading fees. We even run a live A/B experiment on our own execution — two identical books, one paying full fees, one earning maker rebate pricing — to measure exactly what execution costs us.
Chase backtest records
A strategy that returned 300%/yr in a backtest is almost always a bug or an overfit. Our own house rules force us to hunt for the flaw before publishing any headline number.
Hold directional risk
No net-long crypto exposure. No "conviction trades". The book is rebalanced daily to stay dollar-neutral, with liquidity guards that exclude coins too thin to exit cleanly.
Hide the losers
Failed experiments are published alongside the wins — including a feature set we spent weeks building that failed its promotion gate and was rejected. Negative results keep the process honest.
Grid A, B, C — a funnel that
earns the right to trade.
Most trading strategies fail because someone tried a hundred ideas and kept the luckiest one. Our answer is a three-stage elimination funnel with the rules locked in advance. Thousands of candidate configurations go in the top; at most one strategy family comes out the bottom — and even then it must survive weeks of simulated trading before real money is at risk.
Think of it as three rounds of a talent show. Round A auditions the singers (which prediction model can actually rank coins?). Round B picks the song and the outfit (how big should the portfolio be, how often should it change?). Round C is the live final in front of a fresh audience (data the model has never seen, replayed 15 times with different random starts). Only an act that wins all three rounds gets on stage with real money.
Grid A — find the signal · which model can rank coins at all?
We trained three families of models — a simple linear baseline (ridge), a price-level predictor (regression), and a ranking model (LightGBM trained to order coins rather than predict prices) — across multiple horizons, each run five times with different random seeds so a lucky run can't win. Every candidate was scored on the same yardstick: the net Sharpe ratio of the long/short book it would have produced, after 4 bps costs per trade.
Animated illustration — every few seconds the three models guess tomorrow's ordering of the same 10 coins, then reality is revealed: green = called within one place of where the coin actually finished, red = missed. The ranker wins because it's trained on the ordering itself — the exact thing the strategy trades.
Honesty check: a macro-economic feature pack we later built failed its pre-registered promotion test here (Sharpe fell to +1.32) and was rejected. The funnel cuts both ways.
Grid B — build the portfolio · how should the signal become positions?
The winning model from Grid A was frozen. Then we swept every sensible construction recipe: hold the top/bottom 6, 8, 12 or 16 coins? Weight them equally or by conviction? Add "stickiness" so positions don't churn on tiny ranking changes? Smooth the signal over 3 days? Each recipe was priced at three cost levels (0, 4, 8 bps) so nothing wins by ignoring fees.
Grid C — the honest exam · does it survive data it has never seen?
The cruellest test we know: a causal walk-forward. The model is retrained each month using only data available at that moment, then trades the following month blind — exactly as it would live. 13 months, 394 trading days, repeated with 15 different random seeds so a fluke can't pass. The pass rule was locked a month before the test ran: at least 70% of seeds must be profitable.
Animated illustration of the real test: each gold cell is a month the model trades without ever having seen it; once traded, it joins the green training pile and the model retrains. The curve is the shape of the actual median-seed equity path — including the −14.5% drawdown it had to climb out of. This exact exam was repeated 15 times with different random seeds: 15 of 15 finished profitable.
Paper trading · real prices, fake money
The two surviving books (a "plain top-8" and a "smoothed top-12") trade live prices every day in a full simulation — same data feeds, same rebalance clock, same cost model — building a public track record before any real capital moves.
Live capital · the only judge that can't be fooled
On 6 Jul 2026 the plain top-8 book went live on Hyperliquid mainnet with real funds — deployed as two identical twins that differ only in how they execute orders (see the execution experiment in the results section). Every fill, fee and P&L tick is recorded and published on the live dashboard.
Inside the K-of-N book:
8 longs, 8 shorts, zero market bet.
"K-of-N" simply means: from a universe of N = 50 coins, hold the best K = 8 long and the worst K = 8 short. Here is one full day of the machine, end to end.
Net market exposure ≈ $0 at every rebalance. The book earns the spread between its longs and shorts.
Rank
An ensemble of 15 LightGBM ranking models (one per seed) scores all 50 coins on ~24h relative performance. The median score across seeds is the signal — no single model can dominate.
Filter
A liquidity guard drops any coin whose trailing 30-day median daily volume is below $2M — we never hold what we can't exit. Delisted or flagged coins are excluded automatically.
Trade
Target positions are computed; only the difference from yesterday is traded. Typical daily turnover is a fraction of the book, keeping fees low.
Monitor
Hourly drift checks, per-coin loss guards, a kill-file for manual halt, and minute-level equity snapshots feeding the public dashboard. Everything is logged to an audit database.
Crypto's problem for investors has never been a lack of upside — it's the −70% winters. A dollar-neutral book steps out of that fight entirely: it can make money in a crash, a rally, or a flat chop, because it only needs the gap between strong and weak coins to persist. The price of that comfort is a lower ceiling — this compounds like a bond-plus strategy, not like holding Bitcoin in 2021.
What you're paid for
- An uncorrelated return stream. The backtest book's daily returns are driven by cross-sectional dispersion, not market direction — a genuine diversifier next to equities or a long-crypto bag.
- ~14%/yr net in the honest backtest. The causal walk-forward compounded +15.3% over 394 days net of 4 bps/leg costs (≈ +14% annualised) — with every seed of 15 finishing profitable.
- All-weather mechanics. Short legs profit in drawdowns; the book made money across both chronological halves of the test window, including choppy regimes.
- Fee tailwind still to harvest. Live maker-execution is already cutting fees ~62% vs taker on the twin book; if the 6-week A/B confirms, the whole book moves to the cheaper mode.
- Capacity headroom. At today's size we are invisible to the market. The $2M/day liquidity floor supports a book in the low hundreds of thousands before slippage bites.
What can hurt you
- Drawdowns are certain. The backtest's worst peak-to-trough was −14.5%; seeds ranged −11.6% to −15.7%. Expect stretches of months underwater. If that's intolerable, this isn't for you.
- Model decay. Whatever pattern the ranker exploits can fade. Mitigation: monthly retraining, 15-seed ensembles, and pre-registered kill criteria — but decay risk never goes to zero.
- Live sample is tiny. Five days of live trading proves the plumbing, not the edge. Statistically, the backtest is still the best estimate; live confirmation takes months.
- Exchange & counterparty risk. All capital sits on Hyperliquid. An exchange failure, hack or delisting cascade could impair capital regardless of strategy performance. (A TON delisting has already been handled cleanly, but the risk is structural.)
- Crowding & correlation spikes. In violent liquidation cascades, longs and shorts can briefly move together against the book; funding costs on shorts can bite. Position caps and the liquidity guard limit but don't eliminate this.
- Operational risk. One developer, one stack. Mitigations: kill-files, per-coin guards, full audit logging, and infrastructure spend in this raise — but key-person risk is real.
Backtest, paper, live —
the same strategy at three altitudes.
Three views of one book. The backtest is the longest and most statistically meaningful; paper and live are short but they are real-time, un-cheatable confirmations that the machine does what the research says it does.
- Annualised (compound)
- ≈ +14.1%/yr
- Sharpe (median replay)
- +0.80
- Sharpe (plain top-8 book)
- +1.30
- Avg edge per day
- +4.09 bps
- Max drawdown
- −14.5%
- Seeds profitable
- 15 / 15
- Peak so far
- +3.7%
- Max drawdown
- −1.3 pts
- Sister book (smoothed top-12)
- +2.9%
- Rebalances completed
- daily, 0 missed
- Annualised run-rate*
- extreme — see below
- Account equity
- $473
- Max drawdown
- −1.7%
- Fills executed
- 109
- Missed rebalances
- 0
- Annualised run-rate*
- ≈ +61%/yr
Annualised figures — and which one to believe
You asked for annualised numbers; here they all are. But annualising a few days of returns produces silly numbers, so we show the sample size next to each and tell you plainly which figure we would anchor on.
| Track | Raw return | Period | Annualised | Should you trust it? |
|---|---|---|---|---|
| Backtest (causal walk-forward) | +15.3% | 394 days | ≈ +14.1%/yr | ✓ Anchor here.13 months of genuinely unseen data, 15 seeds, net of costs, published pass rules. This is our honest central estimate. |
| Paper book | +2.9% | 8 days | ≈ +248%/yr | ⚠ A hot streak, annualised.Eight days is noise. We publish it because hiding good starts would be as dishonest as hiding bad ones — but nobody should compound eight days. |
| Live account | +0.7% | 5 days | ≈ +61%/yr | ⚠ Too early to mean anything.What five clean days do prove: the pipeline executes, stays dollar-neutral, and tracks its paper twin. The edge itself needs months to confirm. |
All figures as of 11 Jul 2026. Live figures update continuously on the live dashboard.
Twin A — Taker
Half the live book executes the simple way: take whatever price is on the screen. Fast, certain, expensive. 52 fills so far, $0.34 total fees on ~$250 gross.
Twin B — Maker
The other half posts patient limit orders and only crosses the spread as a last resort. Since day two it has converted 100% of its volume at maker rates. Six-week readout: 17 Aug 2026. If P&L confirms the fee saving, the whole book switches — a permanent ~2.8 bps/fill tailwind.
One model family passed the funnel.
Seven more are queued at the top.
Today's live book runs on a single architecture — a gradient-boosted ranking model. The funnel, the data, and the execution stack are architecture-agnostic: we can drop any model into the top of Grid A and let the same locked rules judge it. The next campaign runs seven neural architectures through the identical gauntlet, on the same 50-coin universe. Each survivor is a new, partially-uncorrelated return stream we can blend into stronger K-of-N books.
Different model types "see" different patterns — one reads momentum, another reads shape, another reads seasonality. Two profitable strategies that win at different times combine into one smoother strategy — same return, smaller drawdowns. That blend (an ensemble) is the single cheapest improvement in quantitative finance, and it's what the seed capital funds.
MLP
Dense layers over the same features the ranker uses. Not exotic — but neural nets capture interactions trees miss, and it sets the floor every fancier model must beat.
GRU
Reads each coin's recent history as a sequence rather than a snapshot — natural fit for momentum and mean-reversion patterns that unfold over hours to days.
TCN
Convolutional filters slide across the price history looking for repeating shapes. Trains fast, hard to overfit, strong track record in time-series work.
ModernTCN
Large-kernel convolutions that mix information across time and across features — state-of-the-art results on standard forecasting benchmarks at moderate cost.
PatchTST
The architecture behind ChatGPT, adapted to time series: history is cut into patches and the model learns which patches matter. The strongest published transformer for forecasting.
iTransformer
Inverts the usual design: attention runs across the 50 coins rather than across time — philosophically the perfect match for a cross-sectional ranking strategy.
KAN
Kolmogorov–Arnold Networks learn the shape of each input's effect rather than fixed weights. Unproven in finance; strictly time-boxed. High variance, cheap option.
NLinear
A near-linear model that famously embarrassed complex transformers in academic benchmarks. It runs in every campaign as the honesty control: anything that can't beat it doesn't ship.
First-pass compute: ~£2.9k
Plus seed-count doubling on promising arms, monthly production retrains, and re-runs — the full infrastructure envelope is costed in the ask below.
What happens to a new strategy that passes?
Runs the funnel
Identical Grid A/B/C gauntlet, same locked pass rules. No exceptions, no re-rolls.
Papers ≥ 4 weeks
Trades real prices with fake money until the live-vs-research gap is measured.
Goes live small
Survivors get a modest slice of real capital from the seed pool, tracked per-strategy.
Joins the ensemble
Decorrelated survivors are blended into combined K-of-N books — smoother equity curve.
Or gets retired
Pre-registered kill criteria: sustained underperformance means capital returns to the pool.
£25,000 to scale a machine
that already works.
Not to build something — it's built, tested, and trading. This raise industrialises the research (GPU infrastructure) and gives the next generation of strategies real capital to trade.
£10k — Research infrastructure
- Architecture campaign — first pass, 7 models × Grid A/B/C~£2.9k
- Seed-count doubling & re-runs on promising arms~£2.5k
- 12 months of monthly production retrains~£1.6k
- Data, storage & database replica (3+ yrs tick history)~£1.0k
- Contingency (GPU price moves, failed-run budget)~£2.0k
£15k — Seed for new strategies
- Deployed only to strategies that pass Grid C and ≥4 weeks of papergated
- Per-strategy allocation, tracked & published independentlyaudited
- Survivors blended into ensemble K-of-N bookscompounding
- Pre-registered kill criteria — losers return capital to poolprotected
- Target end-state: 3–5 decorrelated live books12 mo
Why this is the right moment: the expensive part — the data pipeline, the backtest engine, the funnel methodology, the paper/live infrastructure, the execution A/B machinery — is finished and battle-tested by the first strategy's rollout. Every additional architecture now reuses all of it. The marginal cost of a new validated strategy has collapsed from "months of engineering" to "GPU hours plus patience", which is precisely what this raise buys.
Important disclosures
This page is a research program summary prepared for discussion purposes. It is not an offer to sell or a solicitation to buy any security or investment product, and it is not investment advice. Axon Ridge Capital is a private research project, not a regulated fund or investment manager. Backtested performance is hypothetical, has inherent limitations, and does not represent actual trading; live results shown cover a very short period and are not indicative of future returns. Trading cryptocurrency derivatives involves substantial risk of loss, including total loss of capital. Digital-asset venues carry additional counterparty, custody and regulatory risks. Annualised figures derived from periods shorter than one year are extrapolations and are shown for transparency only. Any participation would be subject to separate documentation and applicable law. Figures as of 11 July 2026 unless stated otherwise.