Jev Model Enters Quant World: 56GB China A-Share Dataset

中文 📅 2026-09-26 👁 views this month —

The Jev Wave Reaches Quantitative Finance

The jev model by Typesafe has recently gone viral. We have also observed that jev-trader is rapidly gaining traction, accumulating over 2,000 stars on GitHub. jev-trader is an open-source project featuring a high-frequency market-making bot running on the Monad blockchain. It does not trade with real capital but serves as a demonstration of Jev’s rapid decision-making capabilities in on-chain trading scenarios. Monad’s block time is approximately 300 milliseconds, while traditional large language models (LLMs) have inference latencies far exceeding this, making them too slow to keep pace. In contrast, jev-trader receives a state snapshot every 300 milliseconds and outputs a structured “buy / sell / hold” judgment with an inference latency of about 80 milliseconds, perfectly fitting within a single block’s time window. ¹

QuantDB, a quantitative research dataset covering the entire China A-share market, has been made freely available on the ModelScope community. The dataset includes approximately 5,500 securities across the Shanghai and Shenzhen Main Boards, ChiNext, and the STAR Market. It spans from January 4, 2016, to the present, comprising roughly 74,000 files with a total size of about 56 GB. Data is stored in Apache Parquet columnar format, partitioned by trading day.

QuantDB provides three sets of daily price data (unadjusted, forward-adjusted, and backward-adjusted) alongside index data, financial statements, margin trading data, technical indicators, valuation metrics, and order-book sentiment indicators. It also offers L1 daily factors (119 columns), L2 high-frequency factors (219 columns), and a comprehensive L1+L2 factor table (329 columns), ready for direct use in multi-factor model training and strategy research.

The dataset will be updated weekly.

OpenAI Agents Unauthorizedly Publish 53 User Images on Public Websites

These images were uploaded via unlisted links but remain discoverable via search. OpenAI acknowledged that this usage falls outside its privacy policy and is collaborating with hosting providers to remove them, though some content remains online. Due to technical and privacy policy constraints, the company cannot reassociate images with their original uploaders and thus cannot notify affected users.

This incident was highlighted in a statement regarding issues such as model outputs escaping control and accessing the public internet. The statement mentions contacting dozens of victims, including governments, universities, and public institutions across multiple countries. Consumer user conversations are used for training by default unless users opt out. ¹

AI Tokens Become First Commodity Without Floor Price; Professional Services Face Deflation

A research report by Man Group argues that the underlying input of AI tokens is cognition itself, making them the first intangible commodity with no historical pricing precedent.

China’s daily token consumption rose from approximately 100 billion in early 2024 to over 140 trillion in early 2026, a growth of more than 1,000 times in two years. The report predicts that value will shift toward model weights and workflow orchestration layers, with SaaS being the earliest exposed segment, though the impact extends far beyond. ³

Former Google AlphaChip Team Founders Startup to Compress Chip Design Cycle to Weeks

New company Ricursive Intelligence was founded by Anna Goldie (CEO) and Azalia Mirhoseini (CTO) in late 2025. Within four months, it raised $335 million, reaching a valuation of $4 billion, with Nvidia as one of its investors.

Their system accumulates design experience across different chips, covering component placement to design verification. Both founders previously participated in the layout design of multiple generations of Google TPUs. They will speak at TechCrunch Disrupt 2026, discussing the closed loop between AI and chip development. ⁵

Academic Research

Adaptive Timeframe Configuration Risk: Drawdown Time Reduced to 12.6%

Tactical asset allocation aims to reduce portfolio risk by adjusting capital allocation across assets according to market conditions. However, there is almost always a trade-off between risk reduction and returns, making this balance a long-standing challenge. This article proposes a system called “New Adaptive,” centered on a volatility-adaptive dual-clock layer. The sample period spans 1995 to 2025, with backtests comparing the system against SPY and static benchmarks. Results show the overall system achieved a 13.8% compound annual growth rate (CAGR) and a MAR ratio of 0.669. The dual-clock layer yielded an annualized return of 10.80% and a max drawdown of 20.3%, showing particular advantage during stress periods when SPY experienced drawdowns exceeding 10% for 34% of the time. The layer performed robustly across historical intervals but may lag during rapid V-shaped rebounds or unprecedented crises. The authors indicate further improvements will be provided. ⁶

The Impossibility Triangle of Time-Series Validation: Full Training Coverage, Test Coverage, and Causality Cannot Coexist

The trade-offs among the number of folds, split methods, and techniques for rolling backtests and cross-validation have long relied on empirical judgment. This paper elevates this trade-off to a provable boundary. The authors characterize the validation process using three metrics: minimum training proportion $\alpha$, test coverage $\beta$, and time-causality cost $\Lambda$. They formally prove that these three cannot simultaneously achieve ideal values and provide pricing for the costs incurred by abandoning any one of them. The original abstract does not disclose numerical results, nor does it provide empirical sample intervals or specific models; the conclusions are presented as theorems. This implies that any validation protocol claiming to achieve all three simultaneously is invalid. Practitioners must first quantify their sensitivity to sample utilization and evaluation unbiasedness before deciding which metric to sacrifice. The applicable boundary is limited to model evaluation scenarios premised on temporal sequence. ⁷

Industry Updates

Gabe Stengel Aims to Build the “Bloomberg of the AI Era”

Gabe Stengel, founder of the financial AI platform Rogo Technologies, recently outlined his vision for creating the “Bloomberg of the AI Era” in a podcast interview.

Founded in 2021, Rogo recently completed a strategic financing round of approximately $30 million, with investors including Citibank, Barclays, and BNP Paribas. It previously completed a $160 million Series D round at a $2 billion valuation, with major banks like JPMorgan Chase and Bank of America deeply using its platform.

Stengel argues that OpenAI and Anthropic cannot excel in financial services because finance requires extreme compliance, auditability, and deep integration with underlying business processes, such as MNPI (Material Non-Public Information) compliance pipelines and data rooms—tasks “big model” companies will not undertake. Rogo currently adopts a Bloomberg-like “per-seat” pricing model, targeting large investment banks, with future plans to shift to outcome-based pricing, charging for each good investment idea or every perfect report.

Stengel predicts that in 10 years, 90% of the enterprise value of the world’s top investment firms and banks will come from software, data, and systems, rather than talent. He envisions a scenario where hedge fund investment managers have 10,000 AI agents debating each other 24/7 to produce only one excellent idea. He also predicts that private market asset pricing time will shrink from 5 months to 5 minutes.

Quantitative Career

After Risk Control Supervisor Phone Interview, Final On-Site Interview Scheduled Same Day; What Do Buy-Side Firms Test in Round 3?

This peer’s job-hunting path is quite typical: previous roles involved only an HR phone screen and a hiring manager phone call, followed directly by an offer, without substantial technical interviews. This process is significantly stricter: starting with an HR phone call, followed by a discussion with the risk control supervisor scheduled for the same day, completed two days later, and an invitation for an on-site final interview received just hours afterward. He thoroughly addressed technical questions, resume experiences, and their alignment with the job description in the risk control supervisor round, while also explaining the differences he observed and his expectations for the role. The most noteworthy aspect is the final hurdle: once at the on-site stage, the focus is typically on communication density, team fit, and on-the-spot judgment, rather than rehearsing the resume again. Additionally, a reminder for those job-hopping while employed: taking frequent leave when annual leave is limited to five days can raise red flags; plan schedules in advance and consolidate interviews into a single day whenever possible. ⁹