Vnpy 4.5.0 Upgrade & DSTNet: Outperforming Random Walk
Today’s Highlights
vnpy 4.5.0: Major Backtesting Alpha Upgrade
Vnpy released version 4.5.0 on October 6, following its last update in April this year.
Key updates include: CTA backtests now save and restore by strategy group; Alpha calculations changed from cs_rank to percentile ranking; performance analysis reconstructed; new fill_price and VWAP matching added; ts function accelerated; polygon renamed to massive for US stock K-lines; updates to Mini/TAP/IB/CTP interfaces. Fixes for ranking, indicator chaining, RPC deadlocks, and CTP deadlocks.
Vnpy is a well-known open-source quantitative framework for Trading & Backtesting, featuring trading interfaces, CTA backtesting, Alpha strategies, and portfolio strategies. It has garnered 45k stars on GitHub.1
Private Credit Faces Concentrated Short Selling; Net Shorts Exceed $80M in Early September
Goldman Sachs, JPMorgan Chase, and Bank of America are offering clients a basket of two-way trading instruments encompassing alternative asset managers, Business Development Companies (BDCs), and financial institutions involved in related businesses. These trades are executed through listed vehicles and proxy targets, covering BDCs, CLOs, and alternative asset managers. The sample ETF shares increased from 2.9 million at the end of August to 6.4 million in mid-September, with their proportion of asset management scale rising from 2.1% in August to 2.7% in mid-September, up from 3.4% in March. 2
Hedge Funds Increase Short Positions on Euro; EUR/GBP Hits 16-Month Low
Morgan Stanley advised selling EUR/AUD and EUR/CHF to address fiscal and political risks, as well as a more dovish policy outlook amid bond volatility. The EUR/USD briefly fell 0.7% to 1.1176, nearing a one-year weak level against the JPY. Hedging costs for EUR/JPY volatility exceeded those for USD/JPY by over 100 basis points. Interest rate swaps indicate expectations for rate hikes have dropped from four 25-basis-point hikes to three by September 2027. 4
Big Tech Updates
Morgan Stanley Non-Dev Staff Face Hurdles Gaining Code Tool Access
The Morgan Stanley FID sales team plans to launch operational automation involving tools like VSCode, Python, Git, GitHub, and Claude Code. Internal consensus holds that obtaining necessary development authorizations is extremely difficult. They are currently seeking specific pathways and timelines from those with relevant experience. 6
Quantitative Research
DSTNet: Causal Multi-Horizon Financial Forecasting Network, Surpassing Random Walk
A latest paper from arXiv.
Wavelet-based financial predictors typically use transformations only for denoising or reduce them to a single spectral snapshot at the prediction origin. Consequently, convolution of generated coefficients is often bilateral, prone to look-ahead bias.
DSTNet differs by preserving the recent evolution of filter bank amplitudes as a causal dynamic spectral trajectory. This trajectory is constructed using seven lagging technical indicators over a 20-day backtest window, employing a one-sided Morlet-derived filter bank, with explicit warm-up for left-boundary transients.
A factorized scale-time spectral Transformer performs attention calculations along both the time axis and the filter bank axis. A learned gating mechanism fuses the spectral branch with a CNN-BiLSTM, while specific prediction-step gating outputs forecasts for one, three, five, and ten days in a single forward pass.
We evaluated seven stock indices and gold under a unified expanding window protocol and an untouched one-year holdout set, comparing against nine learning baselines and a random walk persistence benchmark. Under MAE and MAPE, persistence was the strongest among ten fixed competitors in 29 of 32 sequence-step units. DSTNet was the only model outperforming it in every unit: leading by 0.7% to 0.9% on the one-day step, and by 3.4% to 4.5% on the ten-day step. On the one-day step, paired tests support DSTNet’s superiority over weaker learning baselines, though results remain inconclusive relative to persistence and the strongest learning predictor.7
Portfolio Optimization Minimizing Worst-Case Conditional Risk Under Reward-Penalty Mechanisms
When the joint distribution of risky asset losses is unknown but known to belong to a class of multivariate sets, balancing robustness and incentive compatibility in tail risk measures becomes challenging. To address this, a tail mean-loss framework with incentive and constraint terms is introduced. Assuming the loss vector’s distribution lies within an uncertain set, the decision target is to solve for capital weights using the most conservative valuation on the set. Due to truncated text, available conclusions are limited to the above modeling setup and objective description. No quantitative results such as backtest intervals, returns, Sharpe ratios, errors, or significance levels are disclosed, preventing further quantification of improvement magnitude. 9
Quant Career
Quant Job Express (2026-10-09): 28 New Roles at 10 Companies
[This Week’s Selection] G-Research · Linux Platform Engineer (London, GB)
G-Research’s London headquarters is hiring a Linux Platform Engineer to build and maintain core systems supporting quantitative research. Responsibilities include image building and release, automated provisioning, package management tool development, Ansible-as-code deployment, vulnerability response and hardening, and observability construction, while exploring autonomous agent operations.
Requirements include solid Linux practical experience and deep knowledge of kernel startup files, proficiency in Python and automation tools, familiarity with security hardening and compliance, and performance tuning experience is preferred.
The team gathers top-tier research and engineering talent, offering competitive salaries plus bonuses, 30 days of annual leave, and full relocation support, with strong growth potential.
Since Quant Loop launched the Job Express last week, we have identified 28 new roles across 10 companies.
City distribution: 6 in the US (46%), 5 in the UK (38%), 2 in China (15%). The US has the most postings (with 15 roles lacking city labels).
Common requirements: All demand solid professional backgrounds and cross-team collaboration skills to jointly support quantitative investment and research innovation. 12
Word count: 2664 / 2400