The Finance Gazette

Nov 9 – Nov 15, 2026 · Combination is not magic. Count the costs, test the skeptics, and read the easy paper first.

Boolean Logic Shows Why Fancy Factor Combos Cost So Much

Factor investors love to stack signals, momentum plus low volatility, betting the combo beats any single screen. A Walsh-style breakdown splits portfolio returns into exposure to each signal and to the interaction between them, then shows interactions decay faster than the signals that built them. Threshold rules like top-decile picks also throw away rank information that smooth combinations keep, so they trade more for the same interaction. Tested on U.S. equities from 1973 to 2024, the turnover math lines up almost exactly (slope 1.02), and costs bite as predicted: 58% of single-signal portfolios clear their spread costs, but only 6% of three-way interaction portfolios do.

The Logic of Factor Portfolios: Boolean Rules, Interactions, and Trading Costs · Ziyao Wang, Svetlozar T. Rachev, Frank J. Fabozzi · arXiv:2611.09834

An AI That's Not Allowed to Lie to Itself About Backtests

Most 'AI discovers alpha' stories skip the part where the backtest is quietly overfitting wearing a nice Sharpe ratio. This loop, called Autonomous Quant Researcher, lets a language model propose strategies only as strict JSON, never code, then runs them through chronological discovery, validation and a one-shot holdout that can't be reopened. Acceptance needs every gate to pass at once: positive net P&L after costs, a bootstrap edge, survival at higher fees, yearly stability. Across two dozen campaigns spanning 0DTE options, overnight SPX verticals and Kalshi crypto binaries, most hypotheses simply failed and got reported as failures. The best cautionary tale: a survivor with a Sharpe near 3 in validation that collapsed to 0.26 the one time the holdout was used. A Hacker News post flagged the paper but drew zero comments, so the thread proves nothing either way.

Hard to Fool: A Bounded LLM Research Loop for Quantitative Finance, and What Two Dozen Campaigns Found · Henry Zhang · arXiv:2611.06607

One Risk Model Isn't Enough, So This One Asks a Committee

Diversifying across stocks doesn't help much if every risk estimate comes from the same flawed model. This framework adds a second layer of diversification across risk models themselves, pooling several loss distributions with an exponential aggregation function that leans harder on the big losses. The flagship version, DRP expected shortfall, weights each model by how its risk perception compares to the others before combining them. Backtested strategies built on DRP-ES post higher Sharpe ratios and shallower drawdowns than market benchmarks across U.S., Chinese A-share and South Korean markets. Compared with portfolios that only diversify across assets under one model, the DRP approach stays more robust when conditions shift.

Diversified-Risk-Perception Risk Measures and Portfolio Selection · Yang Liu, Yuhao Liu, Jiani Shen · arXiv:2611.11803

Letting a Neural Net Pick Which 5-Minute Bars Matter

Realized volatility measures usually average intraday price moves with the same blind uniform rule, whether that five-minute bar was informative or not. Neural Realized Measures instead uses a neural network to learn positive, mean-normalized weights for each intraday chunk directly from the forecasting objective, while keeping the raw realized measure as a special case when weights go uniform. In Heston-model simulations the learned weighting cuts RMSE for both same-day and next-day variance forecasts across every tested regime. On six cryptocurrencies at five-minute resolution it improves QLIKE in all 54 paired comparisons, with average gains between 7.23% and 9.22% depending on the forecasting model used. SHAP analysis on the learned weights turns up systematic time-of-day patterns, hinting at which parts of the trading day carry real signal.

Neural Realized Measures: Forecast-Supervised Intraday Aggregation for Volatility Forecasting · Brian Sing Fan Chan, Jennifer So Kuen Chan · arXiv:2611.07287

Diversify Your Models Like You'd Diversify a Portfolio

Stock-return prediction has the same problem fund managers complain about: pick one model and you're stuck with its blind spots. This ensemble framework bags linear predictors together with nonlinear ones so each acts as a specialist whose errors don't correlate much with the others, then leans on a bias-variance-covariance decomposition to show this mirrors ordinary portfolio diversification. On CSI300 daily data using the Alpha158 factor set, the mixed ensemble lifts Rank IC by more than 13% over standalone neural networks and edges out tree-based baselines including XGBoost and Double Ensemble. The result reads like a vote for breadth over depth when it comes to model choice, though single-market testing leaves open how far the gains travel.

Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction · Danning Li · arXiv:2611.12389

AI Hedger Beats Delta Hedging, Except When Stocks Actually Rise

A market maker's hedging bot that was raised on real history instead of a tidy simulated market turns out to hedge differently than one trained on theory. The paper trains a distributional reinforcement learning (RL) agent on S&P 500 options data from 1996 to 2017 and tests it out of sample from 2018 to 2023, where it trims tail losses by 3.1% against delta hedging and up to 8.9% in falling markets, but gives back 7 to 10% when markets rally. The authors trace this to a state-dependent short net delta baked in by exposure to real crashes, a tilt that vanishes when the same agent trains on geometric Brownian motion instead. The paper cautions that the robustness claims rest on a single test path, so take them as provisional.

Risk-Sensitive Reinforcement Learning for an Options Market Maker: An Empirical Study with Real S&P 500 Options · Jacky Chen, Ing-Haw Cheng, John Hull et al. · arXiv:2611.10956

Central Bank Scenario Tool Puts Hormuz Shock at 4.8% Inflation

Central bank scenarios usually rely on linear models, but this paper turns the scenario into a full probability distribution instead of one point guess. The authors combine an expert-drawn tail path for oil inventories with a nonparametric model of the oil market and nonlinear Phillips curves for energy, food, and core PCE inflation, assembled as a flexible model specification rather than a single linear equation. Run with only information available in February 2026, the simulated shortage pushes headline PCE inflation to a peak near 4.8% by late 2026, with household expectations climbing about three percentage points and inflation still likely above the Fed's 2% target two years later. The paper presents the scenario as a probabilistic exercise.

Fueling Inflation Risks: An Oil Shortage Scenario · Christiane Baumeister, Todd Clark, Florian Huber · arXiv:2611.12347

Portfolio AI Learns to Say 'No Thanks' to Its Own Predictions

Teaching a portfolio to doubt its own crystal ball sounds like a recipe for paralysis, but this paper builds that doubt into the plumbing instead of papering over it. Its controller chooses among Safe, Follow, and Reverse actions for each signal, with an evidence gate that can abstain by shifting unsupported directional bets back to Safe, plus a delayed collateral budget that reserves loss capacity before any exposure gets funded. On a 24-stock development set the overlay tracks equal-weight performance at a fraction of the turnover, but on a frozen 82-stock holdout the pre-specified Reverse orientation fails to beat equal weight, and simply flipping to Follow afterward does not recover the edge either. The paper states that the contribution is safety, not demonstrated alpha. The paper does not claim demonstrated alpha.

Safe Portfolio Learning from Untrusted Predictions: Signed Trust, Abstention, and Cross-Universe Polarity Failure · Amirhossein MohammadpourAzari, Mohammad Zarei, MohammadAmin Fazli · arXiv:2611.04826

VIX Model Remembers Old Volatility Longer Than Markov Models Do

Volatility, it turns out, holds a grudge, and this paper tries to price that grudge directly instead of assuming the market forgets every tick. The rough instantaneous squared VIX model swaps the usual Markovian short memory for fractional memory governed by Mittag-Leffler decay and derives a semi-closed-form pricing formula for VIX options via an affine Fourier-Laplace transform. Calibrated on daily VIX term-structure data and option quotes from January 2018 to June 2025, the fractional-memory version cuts out-of-sample pricing error by 13.06% in MSE and 6.33% in MAE versus the Markovian benchmark, with larger gains at short maturities and high-VIX states. The paper also reports improvements over several other recent benchmarks, though these are the authors' reported comparisons.

A Rough Instantaneous Squared VIX Model for Term Structure and Option Pricing · Tianfeng Yang, Wen Su, Blanka Horvath · arXiv:2611.07120

Sector Rotation Strategy Stops Flip-Flopping, Drawdown Shrinks to 14%

Whipsaw trading strategies are the financial equivalent of changing lanes every five seconds, and this paper tries to make the model commit to a lane for a while longer. Built on nested clustered optimization, the framework replaces the usual symmetric covariance matrix with a downside-only semi-covariance measure, so upside momentum in soaring sectors is not penalized the same as actual losses, and layers a sticky hidden Markov model on top to resist constant regime switching. Across 65 out-of-sample months from August 2020 to December 2025 the approach lifts CAGR from 10.65% to 13.41%, trims maximum drawdown to 14.22% against 23.93% for the S&P 500, and cuts regime switches from 16 to 4 and drops monthly turnover from 60.78% to 30.48%. The sticky regime layer, in other words, gave the strategy a calmer disposition. The paper reports these as its own backtest results.

Mitigating Return Drag and Regime Whipsaws in Sector Rotation: An Asymmetric Downside-Aware Nested Clustered Optimization Framework · Supriadi · arXiv:2611.05310

All Gazettes · Finance & Markets Research Club