Daily Summary for 2026-11-03

daily

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: It's the third of November, twenty twenty-six, and this is the day's research.

Warren: 3 new papers came out today.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We'll take the day in one pass, then pull out the papers we're staying with.

Warren: Start us off with A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting.

The day's research: Ruth: The controlled comparison involved nine distinct deep learning architectures using a five-stage benchmark across twelve asset types.

Warren: That testing separates architecture from tuning or random starting points for a clean performance picture.

Iris: ModernTCN and PatchTST emerged as leaders, with ModernTCN achieving a mean rank of 1.333.

Ruth: It also set the lowest root mean square error in eighteen out of twenty-four tested asset-horizon points.

Warren: These two models are statistically superior under those specific testing conditions, supported by significant differences between them.

Iris: Following that, admissible portfolio optimization moves beyond simple mean-variance optimization by introducing admissibility.

Ruth: This framework dictates which information pieces are eligible for use in building a portfolio, prioritizing causal identification.

Warren: The selection process ranks candidate representations based on separator defect and information complexity before constructing the final conditional efficient set.

Iris: That ensures portfolio decisions align with declared claims about assets rather than just optimizing weights after an arbitrary representation is fixed.

Ruth: Finally, there is the implementation of partially active automated market makers exploring how dividing liquidity reduces adverse selection costs while managing tracking error.

Warren: A tailored activeness level can be optimal depending on the specific market dynamics for these systems.

Iris: A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting: Evidence from 918 Experiments Reliable comparison of deep learning architectures for financial forecasting requires controlled evaluation that separates architectural effects from hyperparameter tuning, random initialization, forecast horizon, and asset characteristics.

Ruth: Admissible Portfolio Optimization: Information Constraints, Conditional Efficient Frontiers and the Price of Causal Identification This paper introduces a framework to select the best information class for portfolio optimization based on admissibility constraints.

Warren: Partially Active Automated Market Makers We introduce a new type of automated market maker that divides its reserves into active and passive parts to reduce trading costs while balancing tracking error.

Iris: A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting Reliable comparison of deep learning architectures for financial forecasting requires controlled evaluation that separates architectural effects from hyperparameter tuning, random initialization, forecast horizon, and asset characteristics.

Ruth: Admissible Portfolio Optimization: Information Constraints, Conditional Efficient Frontiers and the Price of Causal Identification This paper introduces a framework to select the best information class for portfolio optimization based on admissibility constraints.

Warren: Partially Active Automated Market Makers We introduce a new type of automated market maker that divides its reserves into active and passive parts to reduce trading costs while balancing tracking error.

Iris: A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting Reliable comparison of deep learning architectures for financial forecasting requires controlled evaluation that separates architectural effects from hyperparameter tuning, random initialization, forecast horizon, and asset characteristics.

Ruth: Admissible Portfolio Optimization: Information Constraints, Conditional Efficient Frontiers and the Price of Causal Identification This paper introduces a framework to select the best information class for portfolio optimization based on admissibility constraints.

Warren: Partially Active Automated Market Makers We introduce a new type of automated market maker that divides its reserves into active and passive parts to reduce trading costs while balancing tracking error.

Iris: A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting Reliable comparison of deep learning architectures for financial forecasting requires controlled evaluation that separates architectural effects from hyperparameter tuning, random initialization, forecast horizon, and asset characteristics.

Lucky paper: 2603.16886: Ruth: So, we’re looking at the controlled comparison of deep learning architectures for financial forecasting today.

Warren: The study used nine distinct models across four families, and it ran nine hundred eighteen independent runs following a five-stage protocol.

Iris: That protocol involves fixed seed Bayesian hyperparameter optimization, freezing hyperparameters based on those results, then training across three random seeds to check robustness.

Ruth: The input features were just the five raw market features: Open, High, Low, Close and Volume.

Warren: The primary loss function they used throughout was the Mean Squared Error or MSE.

Iris: In terms of key findings from this research on A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting, ModernTCN emerged as the top performer with a mean rank of one point three three three.

Ruth: And it set the lowest root mean square error in eighteen out of twenty-four tested asset horizon points, which is pretty strong.

Warren: It also showed that the top-tier models, ModernTCN and PatchTST, maintain their superiority across both short term and long term horizons.

Iris: That stability across horizons is what they confirmed with Spearman rank correlations and the Stouffer combined statistic.

Ruth: What about how much variance this architectural choice actually explains?

Warren: The paper says architecture choice explains ninety-nine point nine zero percent of the raw RMSE variance, while seed randomness only accounts for at most zero point zero one percent.

Iris: That really hammers home that the model structure is overwhelmingly dominant in performance variation compared to initialization variance.

Ruth: And there's this point about complexity versus performance, because it wasn't just bigger models winning.

Warren: They found a non-monotonic relationship where DLinear with about one thousand parameters outperformed Autoformer with about 438K parameters.

Iris: So the specific architectural inductive biases are more valuable than just having a massive parameter count, which they support with H4.

Ruth: That makes sense for someone who's actually trying to select a model for production.

Warren: They also looked at directional accuracy, and it remained exactly fifty percent random across all model and horizon combinations tested.

Iris: So the MSE loss function isn't capturing any inherent directional skill at the hourly resolution they tested, which is a limitation they pointed out.

Ruth: That means if you want to predict direction, you need to redesign the loss function beyond just MSE, according to their conclusion.

Warren: The paper concludes that adopting ModernTCN as the default architecture is recommended because of its general purpose temporal modeling capabilities.

Iris: They also noted that for practical implementation, large kernel temporal convolutions and patch based Transformers consistently beat other alternatives in this area.

More episodes

← Home