Research
We test our data before we sell it. These are the studies behind our design decisions, written up with their methods, their numbers and their limits.
Extended fields and short-horizon direction on Hyperliquid perpetuals
The same gradient-boosted model, trained twice on January to June 2026 and tested on July to September. Adding our extended fields to raw trades and top-of-book data raised the out-of-sample directional hit rate by 3.3 points, with the gain present for every asset, month and horizon tested.
Daniel Garcia and Maya Okafor · 11 min read
Read the studyBTC, ETH and SOL perpetuals on Hyperliquid, predicting the sign of the mid-price return over the next 500 ms, sampled every 100 ms. Trained Jan–Jun 2026, tested Jul–Sep 2026, walk-forward with no overlap. Same model and hyperparameters; only the feature set differs. Results exclude fees and latency, may not generalize, and are not investment advice.
All research
Studies run on our own Hyperliquid capture and, where a study says so, on short measurements of public venue data, with other venues anonymized. Each one states its data window, its method and what it does not show.
How we run a study
The same rules apply to every result we publish, including the ones that make our data look less impressive.
Time only moves forward
Models are trained on earlier data and tested on later data, with no overlap between the two. Tuning happens inside the training window, never on the test period.
Headlines are out of sample
Every number in a summary comes from data the model never saw. Features are frozen before the test period starts, and in-sample figures are labeled as such.
Before and after costs
Where a result implies trading, we report it gross and net of fees, spread and latency. Where it does not, we state the cost hurdle a strategy would have to clear.
Same rows you will get
Studies read from the production pipeline, on the same clock and schema customers receive. Configurations, seeds and per-day results are kept for every run.
Research is published for information only. It describes historical data, may not generalize, and is not investment advice.
Run the studies yourself.
Pre-registered partners get early access to the data behind our research, pre-sale pricing, and a direct line to the people who wrote it.
EARLY ACCESS · PRE-SALE PRICINGPre-register for early access → DANIEL GARCIA · CEO & FOUNDERdaniel@financedatacloud.com →