What “research on Hyperliquid” actually means
Research is the work that happens before a strategy is a strategy. You take a hypothesis — “assets paying high funding tend to underperform over the next four hours” — write a factor that encodes it, compute the factor across the HL universe, evaluate whether the cross-section actually has the signal you think it does, and decide whether to keep researching, drop it, or wire it into a deployable book. Most of the calendar in a quant's year is research, not deployment.
On Hyperliquid specifically, research has different inputs than on an equity desk. You have ~220 perp markets with a wide range of listing history. You have hourly funding as a first-class data series, not a footnote. You have a regime structure that flips between extreme-funding, dispersed-funding, and compressed-funding periods within months. And you have venue-level transparency that lets you replay your own fills against the public state. The research toolkit has to make all of that addressable.
The signal-evaluation toolkit
Keel ships the same evaluators used internally to triage which signals are worth promoting. They wire into the same pipeline that produces the signal — no separate research notebook, no subtly different alignment.
| Diagnostic | What it tells you |
|---|---|
| Information coefficient (IC) | Per-bar cross-sectional rank correlation between the factor and forward returns. Spearman by default — robust to fat tails. |
| Rolling IC | Mean IC over a sliding window. Reveals whether the factor is stable, decaying, or regime-conditional. |
| IC stability (ICIR-equivalent) | Mean IC divided by standard deviation of IC. The IC-Sharpe — a high-IC factor with unstable ICIR is just lucky. |
| Half-life | How many bars until the IC decays to half its peak value. Sets the natural rebalance horizon. |
| Decay curve | IC at each forward horizon (1 bar, 4 bars, 16 bars, 96 bars). Shows whether you are looking at a fast signal or a slow one. |
| Quantile spread | Forward return of top decile minus bottom decile. Tests for monotonicity — a factor with high IC but no monotone payoff is suspicious. |
| Turnover | Fraction of the cross-section that changes deciles each bar. High turnover means the signal is short-horizon and cost-sensitive. |
| Cost sensitivity | Quantile-spread payoff net of a realistic fee + slippage model. The honest filter — many high-IC signals die when you charge them realistic frictions. |
The point of the toolkit is not that no one else computes IC. It is that Keel computes it on the same component that the backtester and the live broker will run. No reimplementation in pandas after prototyping in numpy. No alignment drift between research and production. Build it once, evaluate it, deploy it.
Regime-aware factor research
A factor that has an IC of 0.05 averaged across the full sample and an IC of 0.15 in extreme-funding regimes and -0.02 in compressed-funding regimes is not really one factor. It is a conditional bet on the regime. Crypto markets shift between regimes faster than equities, and any research process that ignores that structure will keep promoting fragile factors.
Keel ships five regime detectors: FundingLevelRegime (mean absolute funding across the universe), FundingDispersionRegime (cross-sectional standard deviation of funding), plus vol-, trend-, and correlation-based regimes. Each emits a scalar GlobalSeries you can use two ways in research: as a conditioning variable when you stratify IC and rolling-IC plots, and as a gate or scaler in the live pipeline when the research convinces you the regime matters.
The research flow becomes: compute IC unconditional, then compute IC inside each regime bucket, then decide whether the conditional structure is real (and worth a regime gate) or random (and not worth the complexity).
Hyperliquid-specific considerations
- Funding as alpha, not friction. On most equity desks funding is a borrow cost you subtract at the end. On HL it is a tradeable data series that has its own IC, its own decay, and its own regime structure. The research toolkit treats funding as a factor input, not a P&L deduction.
- Regime shifts inside months. Equity regime research often works on quarterly buckets. HL regime structure moves on a weekly-to-monthly cadence. Rolling-IC windows of 50 to 200 bars (12 hours to 2 days at 15-minute resolution) are more informative than the typical 252-day rolling window from an equity context.
- Newer listings have short history. A factor IC computed over a universe that includes assets with 30 days of history is dominated by survivorship and selection effects. Keel's evaluators let you restrict to a min-history universe per evaluation — the honest version of the IC, not the version that looks best in a screenshot.
- Quoting and tick conventions. Cross-sectional evaluations need consistent units. The data loaders standardize price returns, funding rates, and OI deltas across the universe so a cross-asset IC computation is not silently comparing percent moves on one asset and basis-point moves on another.
An end-to-end research flow
A typical session: load the HL universe and a candidate factor as a pipeline component; run the signal evaluators to compute IC, ICIR, decay curve, half-life and quantile spread; stratify the IC by regime to check conditional structure; if the unconditional or conditional IC is convincing, wire the signal into a strategy skeleton with a forecast-weight aggregator and a vol-targeting overlay; backtest with funding decomposition and realistic costs; compare against a benchmark; ship to paper, then live, with the same pipeline definition.
The point of designing the toolkit this way is that you can stay in one tool from hypothesis through to deployed book. The research artifact and the live artifact are the same Python object. That removes the most common source of slippage between backtested IC and live performance — the rewrite that happens when you graduate from notebook to production.
Try it
Open the platform and run a factor against the HL universe with the signal evaluators. Same pipeline runs live when the research checks out.