Reading the book

The book is the terminal's shortlist: the names a daily rebalance would hold right now, ranked by the forecast over the class model's own horizon.

It is built from the same daily artefacts everything else on the screen comes from, and it has three moving parts worth understanding before you use it.

What GBM is

The forecast behind the book is GBM. GBM is a LightGBM gradient boosting model, one per asset class, trained offline on that class's own history. The daily classes train through ops/gbm_train_class.py and crypto through ops/gbm_train.py, on the class's momentum and volatility ladder plus its shape descriptors, and the model returns a forward return estimate. It is a relative strength ranker, not a probability: the useful part is exclusion and flagging, not chasing the top names. Crypto's tails are not a dependable short signal: the published claim that its top decile lost 6.2 percent at 30 days came from a 163 day window, and a later re-measurement on the perp store did not reproduce it, the measured year's 30 session decile median landing about plus 6.9 percent, while equities, by contrast, are a tails market.

The horizon is the model's own, and the column header says which: 24 hours on crypto, five sessions on the daily classes, read from each model's manifest.

Selection

Names qualify on liquidity first, with a floor on 24 hour dollar volume, and then on forecast. The book holds the top decile by cross-sectional percentile, capped at a fixed count. Percentile means strength relative to the field today, not a probability of going up. A name at the 99th percentile is the strongest name on the screen this build, and it can still fall.

Hysteresis, so the book does not churn

Entry and exit use different thresholds. A name enters at the top decile and stays until it falls below the keep threshold, which sits lower. Without that gap the book would reshuffle every hour on noise, and every reshuffle would look like a decision.

The anchor is the day, not the hour

Entries and exits are measured against the last daily anchor. Crypto builds hourly, so the underlying data moves every hour, but the book you are meant to act on changes once a day. The panel shows what the next rebalance would change, not what happened in the last sixty minutes.

What it deliberately ignores

There is a fade list at the bottom of the distribution, and it is informational. The short side of this model failed out of sample, so the product shows it without pretending it is actionable. Model health is reported alongside the book through the standing IC report, so a decaying model is visible rather than quietly left in place.

How to use it in practice

Open the book first, then check the leaderboard for the names you do not recognise, then check the detail panel for shape and coverage. A name with patchy data carries a coverage flag and should be treated as a weaker observation, not a stronger signal. If you want the reasoning in words rather than columns, the chat panel answers from the same book on screen.