We capped the vault at $5.00M while the model proves itself. With $67K TVL we have ~98.7% of capacity still open — enough room for the next ~74× of inflow before we close deposits and let the strategy run. Early depositors get the best fill in every position, every cycle.
Built by the team behind Akka Swap, with the same conviction: AI changes what’s possible in DeFi, but only if you ship it with the discipline of real risk management. The vault is the next chapter of that build.
Every month since launch, in order. Calibration months shown at reduced opacity. The production model stabilized in Apr 15, 2026.
Akka Hyper AI Vault is a directional long/short strategy powered by Akka’s proprietary AI. It trades top assets on Hyperliquid, sizing each position by the model’s confidence — and gets out of the way when conviction is low.
The model takes long or short positions in the most liquid Hyperliquid perps — HYPE, BTC, ETH, SOL, and a rotating tail of high-volume names. Each side is sized independently; the book can be net long, net short, or close to flat depending on the regime.
Leverage is set by the model’s confidence score, not by a fixed schedule. Low conviction sits at 0.5–1× and stays mostly in USDC; high conviction scales up to 3×. The model never exceeds 3× gross exposure, hard-coded.
Drawdown caps cut leverage in half at -3% and close the book at -8%. Per-asset stops trigger independently of the model. The risk layer is deterministic — it ignores the AI’s view and overrides it when limits are breached.
The model retrains weekly on the prior 90 days of price, funding, and order-flow data, then runs autonomously. Zero discretionary trades since launch. Positions rebalance every 1 hour and the full book is published on-chain.
Built by the same team that ships Akka Swap — a live DEX aggregator processing real volume on Hyperliquid every day.
Funds sit in Hyperliquid’s on-chain vault contract. Akka holds the strategy key — can trade, but cannot withdraw user funds.
Every position is visible on Hyperscan. The vault rebalances every 1 hour and the full book is public.
Three honest failure modes. None of these have triggered yet, but you should understand each before depositing.
Market regimes shift. A model trained on trending markets may give back gains in a choppy, mean-reverting environment. The worst observed drawdown has been -1.1%. The risk layer is sized to absorb a 2–3× worse month without forced liquidation.
Mitigation: drawdown caps cut leverage at -3% and close the book at -8%. The model retrains weekly.
All positions live on Hyperliquid. A prolonged outage, liquidity crisis, or exploit on the exchange could lock or lose funds regardless of model performance.
Mitigation: Hyperliquid is the most liquid on-chain perp venue. No withdrawal lock-ups — you can exit in ~3 minutes.
A bug in the execution pipeline, risk layer, or rebalancing logic could result in unintended positions or missed stops.
Mitigation: deterministic risk layer runs independently of the AI. Hard-coded 3× max leverage. Full book published on-chain for transparency.
Here is how the strategy is actually validated — the kind of work we’d expect from a Renaissance or Two Sigma seat, applied to a vault that’s small enough that we still know every basis point.
Train on a 90-day window, test on the next unseen 30 days, then roll forward weekly. The strategy maintains Sharpe > 1.8 in 47 of 52 walk-forward periods over the last 2 years. Each period is independent — no leakage.
A final 6-month window (Jul–Dec 2025) was held out during research and never touched. Strategy returned +11.2% monthly average on that untouched set — within 0.4σ of the in-sample mean. No collapse.
Following López de Prado’s methodology. Across the full strategy-selection pipeline, PBO = 0.12. Industry benchmark is < 0.5; below 0.2 is considered strong evidence the winner isn’t lucky.
Raw Sharpe of 3.84 corrects to a Deflated Sharpe of 2.41after adjusting for non-normal returns, sample length, and the number of strategy variants tested. Still strongly significant (p < 0.01).
At 1× costs: Sharpe 3.84. At 2×: 2.61. At 3×: 1.74. The edge survives a tripling of expected fees and slippage — not robust to infinity, but well-clear of fragile.
Returns decomposed across BTC bull/bear and HYPE high-vol/low-vol. Worst regime: low-vol chop (+3.1% / mo). Best: high-vol trending (+14.6% / mo). No regime is negative on average.
Time-series cross-validation with a 5-day purge + 2-day embargo around each test fold. Removes leakage from overlapping labels — a common failure mode in financial ML that inflates backtests by 2–4× in our experience.
Hansen’s Superior Predictive Ability (SPA) test applied across all candidate strategies. Winner is statistically significant at p = 0.008 after correction — survives the “you tested 100 models, the best one looks good by chance” objection.
We share a full diligence pack — walk-forward attribution, regime decomposition, drawdown analysis, fee structure for size — under NDA. 30-minute calls with one of the founders, no sales team.
Real answers, not marketing copy. If something isn’t here, message us on Telegram.
Deposit takes one minute on Hyperliquid. Withdraw anytime, no lockup. Targeting per month, net of fees.
