A week of food, drink & art in "communist" China

Order execution slippage over time is a much bigger cost than exchange trading fees. Below is a proprietary dynamic TWAP execution algorithm powered by deep reinforcement learning that optimizes execution and often delivers "negative slippage" vs target execution price. The data

And ... as predicted, LLMs at @the_nof1 mean revert to 0

2 screens is more than enough for quants, if you're focused on building systems instead of looking at market

The infamous paper well-known in the quant trading circle "Virtue of Complexity of Return Prediction" by Bryan Kelly of @AQRCapital may be one of most important and thought provoking peices you may come across in recent years. It's not that it contains any ideas about alpha

Retail fintwit and crypto bros version of AI trading => plotting LLM PnL * Quant funds & algo prop traders facepalm * Brings out deep learning & reinforcement learning systems built and trained on terabytes and years of data, proprietary alpha discovery mechanisms, alpha-decay

Yeah I don't think most people understand that LLM-based trading is a crapshoot. In fact, Deepseek leading the PnL and GPT doing horribly does not say anything about the models' ability to analyze market data nor indicative of future performance on a walk-forward basis. The edge

Having experienced the latest crypto crash makes generating alpha via building trading strategies with a systematic edge & limited exposure all the more compelling vs. degeneracy.

Now, this is without a core strategy on top - once we layer on a deep learning-based or alternate quantitative strategy, the edge is magnified. In this particular case the deep learning-based strategy used to pair with the unsupervised regime achieves a sharpe of >3 with added

We compared two versions of the model 1) a linear multi-factor model 2) a modified version of Wasserstein K-means (unsupervised learning). Frequency (determined by feature time horizon length) was tuned to around 1-2 days per switch. We can see that cumulative return of the

Here's how the benchmark works => the models are designed to output a general real-time classification of either A) uptrend (0) or B) downtrend (1). On switching to uptrend regime state, we assume a long position, and upon switching to downtrend state, we assume short. No added

This is a simple demonstration why market regime models can provide quantitative or ML-based strategy a "base edge" and is far from obselete. Of the various regime types, below is a a benchmark of a trend regime model run on the largest volume Chinese Equity Index Futures (S&P500

Well I'm glad that there's a random Wolfgang Puck Steakhouse across the creek from the Chinese village villa that I'm in this weekend. #ruralchina #nature

There are hundreds of fine details, and hence is probably not for everyone. There seems to be 2 camps of people in whether copious amounts of compute would cause strategies to crash and burn or perform better. I'm in the latter camp since been a practitioner since 2022 with

Now in order to achieve this level of speed and scalability in strategy development and alpha extraction, you definitely need a strong ML/AI pipeline (at least 6-7 junctures) including everything from data/feature processing, training, simulation/search, stress testing,

Live trading started Jan 2024. None of the substrategies are overoptimized. Only volatility and regime-based parameters were adjusted based on asset-level distribution and time horizon, so minimal work was required on our behalf. As you see without the aggregate effect of

A robust deep learning-based system and pipeline for strategy and alpha discovery will enable rapid development, evaluation, and testing of various types of low-correlation substrategies (also can be seen as sources of alpha) to synthesize into one superior main strategy

It will be a simple task for us as I myself have deployed hundreds of AI trading strategies across crypto, fixed income, equities, and commodity futures. The proprietary technology and knowhow my firm brings to the table will be injected directly into AlphaNet AI DEX.

This type of growth requires a strong technology-focused AI quant trading team, along with ample compute resources to scale, both of which my firm Tensor (AI backbone of AlphaNet DEX) has. Scaling effectively means expanding strategies across as many assets and strategy-types as

Note that the growth dynamics would be completely different from a typical exchange - where bigger or more is better. On such a platform, users function as an alpha (profit)-preserving collective. Too fast of user growth may erode edge and alpha generation of the existing base.