Optiver AI Contest Evaluates Real-Time Earnings Reaction
Trending Topics
AKShare Fixes Futures Warehouse Receipt Daily Reports
On September 30, 2026, AKShare released v1.19.1, fixing web parsing for the Shanghai Futures Exchange (SHFE) warehouse receipts daily reports, correcting field mapping for additions and deletions at the Guangzhou Futures Exchange (GFEX), and providing explicit error messages and timeout settings for the Dalian Commodity Exchange’s (DCE) anti-scraping measures. 1
SHFE Nickel Futures Overseas Delivery Month Roll Implemented
Recently, the SHFE nickel futures overseas delivery month roll was successfully implemented, completing the final step toward internationalization. GEM Co. Ltd. locked in prices and hedged risks through dual anchoring on the Shanghai and London markets. The rise of recycled nickel combined with the export of Shanghai prices has helped Chinese nickel companies enhance their global influence. 1
Safe Havens? Not Anymore
The iShares 20+ Year Treasury Bond ETF (TLT) hit a new low yesterday, facing an eighth consecutive day of declines. As a traditional safe-haven asset, long-term government bonds have shown instability. Since the Federal Reserve began raising interest rates in Q1 2022, TLT has recorded its sixth quarter with a decline of at least 10%. 2
Big Tech Updates
Optiver Joint Contest Evaluates AI’s Ability to Interpret Stock Price Reactions to Earnings in Real Time
The core challenge addressed by this evaluation is how to enable AI agents to provide interpretable judgments on stock price reactions simultaneously with the release of earnings announcements and conference calls.
The competition launches on October 12. In collaboration with Optiver, organizers have opened two new datasets to participants to improve the interpretability of agents. The evaluation emphasizes real-time performance, distinguishing itself from post-hoc backtested textual analysis of earnings reports. No baseline scores or error metrics were disclosed; participants must tune their models based on the new data. 15
Quantitative Research
Microstructure Explains the Failure of Short-Term Trend Strategies
A study published in July 2026 indicates that short-term trend-following strategies have failed since 2008–2009, while long-term strategies remain effective.
Using data from the SG CTA Index and 100 futures contracts, the study found this phenomenon to be sudden, speed-dependent, and heterogeneous across assets. The study ruled out factors such as capacity constraints, electronic trading, and dynamic changes in CTA and order flow, confirming that the minimum tick size after volatility standardization as a key variable.
For contracts with small tick sizes, short-term trend returns dropped to zero after 2008, while those with large tick sizes remained stable. The mechanism lies in the fact that trend strategies rely on self-reinforcing feedback loops, where directional trading influences prices through market impact, thereby reinforcing signals.
After 2008, high-frequency trading (HFT) replaced traditional market makers. Their flat inventory limits made them unwilling to absorb large directional orders, withdrawing liquidity. In small-tick contracts, the order book is sparse; after HFT withdrawal, there is insufficient depth for trend traders to execute, breaking the feedback loop and causing short-term trend strategies to fail. 1
What Institutions Say, What Media Report, What Positions Do: Who to Follow in the Three-Part Market?
When institutional statements and media narratives are treated as valuation evidence by machine learning signals, is the party shaping the narrative quietly taking reverse positions?
The study disentangles three observable voices in the market: institutional statements (Say), media repetition (Echo), and revealing positions (Do), characterizing the optimal trading and communication strategies of a knowledgeable institution within a linear-quadratic model framework.
The conclusion shows whether speech is worth following depends on communication costs and the degree of position revelation: when media echoes diverge from actual institutional positions, strategies following textual signals will systematically lose money and need to switch to reverse operations. This framework separates the pricing of "what is said" and "what is done," providing testable failure conditions for textual factors. 4
Two Specialized Paths for Large Language Models to Execute Algorithmic Trading Code Generation
General-purpose language models are not weak at writing code, but they struggle to faithfully translate a natural language strategy into executable program logic within a specialized trading framework, execute it on historical data, generate trades, and stay true to the user’s original intent. This is the most difficult aspect of algorithmic trading code generation.
To achieve this goal, the work compares two complementary model specialization mechanisms, mapping strategy specifications to program logic for specialized trading frameworks, and evaluating both executability and semantic fidelity.
The conclusion points to one thing: general code capabilities are insufficient for this scenario. Whether a model can actually run in a specialized framework and produce trade records is the key differentiator between specialized and general models. Due to the truncation of the original abstract, specific data intervals, model sizes, and backtest metrics are not disclosed. Under this theme, trading code generation is shifting from single-round generation to iterative verification with execution feedback. 5
Optimal Liquidation Strategy in Closing Auctions: Comparing DQN and Projected DDPG
After continuous trading ends and transitions to a closing auction, how should remaining inventory be allocated between selling pace in the limit order book and the auction phase?
The study breaks down the trader’s actions into two steps: first selling in the limit order book, then submitting signed auction plans to adjust remaining positions, connecting the two phases using estimated liquidation price signals and intermediate auction feedback. Strategy learning uses deep Q-networks as a baseline, compared with continuous action methods such as projected deep deterministic policy gradient and twin delayed deep deterministic policy gradient (TD3).
Comparative results show that continuous action projected policy gradient methods are better suited for handling constraints on signed auction plans, while liquidation price signals between the two phases and auction feedback are key variables affecting execution performance. 6
Large Language Model-Driven Small-Cap Stock Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals
This study explores using large language models to extract richer signals from financial news than fixed sentiment dictionaries, and investigates incorporating them into portfolio construction methods.
The study proposes an uncertainty-aware construction framework that decomposes model prediction risks into aleatoric and epistemic components, directly inputting them into the covariance matrix of the portfolio allocator, rather than treating risk as a fixed value or only adjusting expected returns. 9
Quant Career
Are Interns Just Doing Chores? See What Jane Street Interns Are Up To
Jane Street’s latest blog post introduces the work of Jane Street interns.
At Jane Street, interns are thrown into real business problems from day one. Most complete actionable or research-valuable outcomes within a single summer, covering five types of roles.
Machine Learning Research: Rapid Trial-and-Error as True Research
Interns’ work resembles exploratory scientific research, often involving large-scale training from the first week. Kavish Kondap built an event-level generative model for market data using autoregressive diffusion models to explore how continuous market data truly is. Monte Bohde used divergent decoding, distillation, natural language autoencoders, and prompt engineering to study how to mitigate large models’ memory of historical results. Other interns worked on per-order prediction networks for order books, KV cache distillation to accelerate prefilling, and quantization-aware training exploration for Mamba models.
Software Engineering: Directly Improving Production Systems
Arsh Koneru designed an activation checkpointing solution superior to PyTorch, enabling training even under tight memory budgets. Theodor Totev built indexing for the message framework Aria, reducing CPU usage in real-world scenarios by 30%. Sai Konkimalla developed a DSL for CPU SIMD instructions using OxCaml. Others used ILP/SAT solvers to improve scheduling for market data applications or optimized feature engineering libraries.
Linux Engineering and IT: Filling Infrastructure Gaps
Jacob Root created a reporting tool that preserves kernel logs even during disk or network failures. Kian Kasad wrote a FUSE server in OCaml, achieving read speeds of 1.3 GiB/s. Max Ohm replaced scattered Windows configuration scripts with a centralized system and, by running old and new systems in parallel, discovered long-hidden issues in the old system.
Trading Desk Operations and Strategic Products: Close to the Business Frontline
TDOE interns participated in convertible bond trading analysis and exotic option cash flow reconciliation, needing to balance system robustness with the urgent demands of trading. Strategic and product interns embedded themselves in teams, led by full-time mentors, working on projects such as automated Linux kernel upgrades, loan fee modeling, and risk limit management for special dates.