Quant Beckman

First: Reduced model --> Then: Full model

@benjmttt A NNet aproximating noise

@quantbeckman The ones who survived treated the LLM as a signal preprocessor not an alpha generator. Different architecture, different expectations, different results. What was the specific breakdown you saw most?

Academy VS reality. This is how the LLM #trading season has ended.

@TheSignlProcess 💯

The compression framing is the useful part. I think most signals fail not because they lack predictive power in isolation but because they discard information that only becomes relevant when the market structure shifts. A sufficient statistic for a stationary market isn’t sufficient anymore when the regime changes.

📘[QUANT LECTURE] Sufficient statistics and minimal signals📘

@theone34824249 There is always something better

@quantbeckman Is there a better model than equations with just reasoning ???

Mmmm since the study uses an unlevered long-only framework over 1991–2020, part of the strong MPO RP result may reflect a bond-friendly era.

Someone asked me about the classical quant pipeline. Well, here it is…

@MMtrades12 It depends on the models category

@quantbeckman What are some methods to unify all of the sub models internally?

Trading system == Model of models

Extract the risk-neutral density from the implied-vol surface and treat it as a probability density over strikes. Each date becomes a point on a density manifold, where geometric distance measures shape change in skew and tails. Compare today’s density to a historical library to detect regime shifts. Use the distance as a trigger for hedging, sizing, and volatility risk limits.

🟥For this one and more papers with code: quantbeckman.com

The paper says they continue training while in the out-of-sample trading stage to adapt to market dynamics. For me this is a fatal flaw.

You could treat the volume profile as a probability density over time-of-day, rather than raw volumes. Then each day is a point on a density manifold, where shape differences are measured by Fisher–Rao geometry. Compare today’s partial-day shape to historical shapes to infer the closest regime (open-heavy, lunch-dip, close-heavy). Use the match to adapt execution schedules (VWAP-like, or similar).

🟥For this one and more papers with code: quantbeckman.com

The unconstrained strategies come with absurd gross leverage. The paper reports leverage-like absolute position sums around 438 for the unconstrained MS strategy and above 2154 in the bi-sort case, then rescales everything to zero-cost, leverage-two portfolios for implementability

@nieledran 💯

@quantbeckman Accept

Call of duty 🤣

@Praxis_Cap Cool!

@quantbeckman Worked on pc.

If you are a quant or algotrader, join us! No more solo-quants here. Link in the comments 👇🏻