An orthogonal tabular sleeve: STRAT-04b

Abstract

STRAT-04b is a single LightGBM lambdarank model trained on a tabular feature stack, deployed cross-sectionally over a fixed seven-coin inverse cluster. On the Pass-B cascade-strict OOS, it produces an annualised dollar Sharpe of +1.51 at 4 bps per leg, with mean per-bar PnL of +0.195 bps, compound return of +16.92 percent, and a maximum drawdown of just 5.41 percent over 8 041 bars. On Pass-A (the chronological tail), the same model loses badly — −4.24 Sharpe — and we have to deal honestly with that. We deploy it only as part of a portfolio with sleeves that hedge its Pass-A failure mode, never standalone.

1 · The model

LightGBM^[1] with the lambdarank objective^[2]. The model is trained on the same tabular feature stack used across the program — the funding, open-interest, momentum, volatility, and order-book pressure features described in the architecture roster — and produces per-coin scores at each bar. The lambdarank objective optimises the cross-sectional rank at each bar rather than per-coin regression error, which makes it natively suited to a cross-sectional strategy.

The deployed strategy uses these scores to size positions on a fixed seven-coin universe — the inverse cluster: TIA, POPCAT, kPEPE, W, JUP, LINK, TON. These are the coins the program-wide IC pattern analysis^[3] identified as the cleanest short-cluster on this universe. The model's cross-sectional ranks are inverted (most-positive prediction becomes the largest short) and the positions are rebalanced daily.

2 · Per-coin contribution

3 · The Pass-A failure: facing it honestly

The Pass-A chronological tail covers February through April 2026. On that window the inverse cluster simply did not behave like an inverse cluster — the names rallied or chopped sideways with elevated funding, and the persistent short signal that drives most of the Pass-B PnL produced losses. The model has no regime-detection layer; it produces the same kind of cross-sectional rank in any window.

Three responses are honest, in increasing order of intervention:

• Acknowledge it. The model is regime-fragile on its standalone Pass-A. This is not a bug, it is a property. Strategies that are good at one thing are usually fragile on something else.
• Hedge it. Pair the sleeve with a strategy whose Pass-A is strong. The static consensus basket STRAT-48 (Pass A +2.93) is the natural pair: the two together are positive on both passes at the portfolio level.
• Pre-register the kill condition. If 90 days of live Sharpe is < 0, the strategy is debunked and retired. The kill condition lives in the strategy's research file and the kill-switch is wired into the live execution stack.

4 · Distinctness — why we keep it in the portfolio

Despite the Pass-A weakness, STRAT-04b earns its place in the live roster because its IC vector is decisively orthogonal to the rest of the architecture cluster. The distinctness score — one minus the maximum pairwise correlation of its prediction vector against every other arm — is 0.61. The modern-mixer cluster has internal distinctness scores below 0.15.

5 · Cost sensitivity

Unlike the K = 5 / 5 LightGBM ranker discussed in the cost-aware backtesting paper, STRAT-04b is not cost-fragile. Its turnover is modest because it rebalances on a fixed seven-name universe (no new names rotating in), and the cross-sectional ranks within that universe are relatively stable from day to day.

6 · The deploy spec

The deployed configuration:

• Architecture: LightGBM lambdarank, 200 trees, depth 5, learning rate 0.05
• Universe: TIA, POPCAT, kPEPE, W, JUP, LINK, TON
• Rebalance: daily at 14:15 UTC (timed for liquidity)
• Position sizing: cross-sectional rank scaled by pos_scale=25, signs inverted
• Risk gates: per-coin notional cap, leverage limit, daily loss kill, per-strategy drawdown kill
• Falsification gate: 90-day live Sharpe < 0 → debunk and retire

7 · Open questions

• Does a regime overlay fix Pass A? A pilot (STRAT-09) tested several volatility and funding-z regime gates as overlays on STRAT-04b — none of them fixed both passes simultaneously. Open: a more sophisticated regime detector that uses cross-arch consensus disagreement as a signal.
• Does the lambdarank objective vs lambdamart vs pairwise classification matter? Not extensively tested on this universe; on track for the next CPCV sweep.
• Universe drift. The inverse cluster was identified on the Pass-B training window. If the cluster decays — names exit or new names should enter — the strategy needs to be re-fit, not just re-trained.