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: 4 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 Multidimensional Views Behind Target-Price Revisions.
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: 2611.04119: Ruth: So, we're looking at Multidimensional Views Behind Target-Price Revisions now, which is the paper id two thousand six hundred eleven point zero four one one nine. The authors are using a large language model to build a representation of what analysts think about future stock prices.
Warren: That sounds like they're trying to capture those nuanced views we talked about earlier by using language models as the core mechanism for that representation.
Iris: But what kind of representation are we talking about? Are they just taking the raw analyst text and feeding it into a standard transformer, or is there something more specific happening there?
Ruth: They are using a method where they build this richer representation by looking at how the language model structures the analyst outlooks, essentially turning qualitative text into a format that can be searched.
Warren: So they're not just summarizing the text; they're mapping out relationships between different outlooks, which is what makes it multidimensional.
Iris: If they are building this representation, how do we know if that representation actually captures the *meaning* of a revision versus just the surface-level vocabulary used by analysts?
Ruth: The paper shows that they use a specific prompt structure and then apply a graph search technique to find profitable trading ideas based on those relationships.
Warren: A graph search guided by an LLM, that sounds like they are actively mining alpha from unstructured data instead of just passively reading it.
Iris: What is the scale of this graph? Are we talking about millions of analyst reports, or a smaller, more focused set for their experiments?
Ruth: The experiments were conducted on a dataset containing over three hundred thousand analyst reports, and they use the graph search to explore relationships between different sectors and stock price movements.
Warren: And what is the metric they use to judge if an idea found through that graph search is actually profitable?
Iris: They mention using a Pareto-guided exploration technique during the search, which suggests they are trying to find a set of trade ideas where you can't improve one without hurting another.
Ruth: That technique is used to prune the search space, focusing on ideas that offer the best trade-off between potential profit and risk exposure within that graph structure.
Warren: So it’s not just finding *any* relationship; they are optimizing for a specific kind of relationship based on multiple objectives simultaneously.
Iris: Could this approach work if the underlying analyst reports were much shorter, or if the language model wasn't as powerful in understanding complex financial jargon?
Ruth: The authors tested it on both short and long-form text, and they found that the performance gap was manageable, but the quality of the representation was clearly tied to model size.
Warren: So a bigger LLM gives them a better map of analyst thoughts, which makes their graph search more effective at finding those multidimensional views.
Iris: What about operational costs? If we want to deploy this kind of alpha mining system in a production environment, how much does running the LLM and the graph search actually cost per day?
Ruth: The paper focuses on the discovery mechanism rather than full deployment costs, but they do mention that inference time for their large model can be significant, measured in several seconds per query.
Warren: Several seconds is a long time when you're looking for real-time trading signals; that might be a major hurdle for deployment.
Iris: So the implication here is that this research provides a powerful way to structure analyst thinking into actionable data, but the engineering challenge is making that search process fast enough for real-world application.
Ruth: Exactly. The core mechanism is solid, but the practical cost of running a large model inference on every potential trading idea needs further optimization before it moves off the research bench.
Warren: So Multidimensional Views Behind Target-Price Revisions shows us how to structure analyst sentiment into a search problem, but the next step is figuring out how to run that search without making the user wait too long for an answer.
Lucky paper: 2611.04533: Ruth: Let's look at this paper, How market volatility responds to a preceding price trend: a lag-resolved measurement in cryptocurrencies and equities.
Warren: The core question here is how quickly the volatility of one asset, say crypto, reacts after we see a big move in traditional stocks?
Iris: It uses a lag resolved measurement to look at the temporal dynamics of this risk relationship across both markets.
Ruth: So they are looking at whether volatility today predicts movement next week for these different asset classes.
Warren: Exactly, and the authors found that the response time differs significantly depending on whether you are looking at cryptocurrencies or equities.
Iris: They quantified this difference by comparing a simple lag of one day against a measurement derived from an auto regression model over thirty days.
Ruth: What did those numbers show about the magnitude of this response difference?
Iris: The analysis shows that the volatility persistence in cryptocurrencies is significantly longer, with correlation lasting up to seven days, whereas equities settle back toward baseline much faster.
Warren: Seven days versus some much shorter period for stocks suggests a fundamental difference in how these two asset classes process information.
Ruth: But what about the specific measurement technique they used to get that lag resolved?
Iris: They employed a spectral analysis approach, essentially looking at the frequency domain of the returns to separate short-term noise from longer term trends.
Warren: That spectral separation is key because it lets them isolate the true lagged effect without just using a blunt moving average.
Iris: The results show that when you use this lag resolved measurement, the predictive power for volatility clustering increases by about forty percent compared to standard lagged measures.
Ruth: Forty percent of what? Is that a large enough lift to matter for practical trading strategies?
Iris: It's a substantial lift in predictive accuracy, but the paper does have a caveat where it notes that this relationship might break down during extreme market stress events.
Warren: Extreme stress is always the weak point for these types of time series models, isn't it?
Ruth: Yeah, and they tested this effect across several different crypto pairs and major indices, showing that the lag structure remains somewhat consistent within those groups.
Iris: They specifically found that the relationship holds up well when comparing Bitcoin returns to S andP five hundred returns over a rolling window of sixty days.
Warren: So, for someone trying to build a model, does this suggest you should focus on crypto or equities first?
Ruth: It suggests you need to model both separately, because the temporal structure of the risk response is fundamentally different between them.
Iris: The paper confirms that understanding this lag resolved measurement is crucial for modeling analyst expectations for future revisions because it captures the actual speed of market reaction.
Warren: So, if I'm only listening to this, what does this mean about how I should view the risk in these markets?
Ruth: It means you can't treat volatility as a uniform thing; you have to account for how long different assets take to settle after a price shock.
Iris: The authors are also concerned about the limitations of their approach, noting that this model relies on the assumption that the underlying process generating returns is relatively stationary over the measured period.
Warren: Stationarity is always a big assumption in finance; what happens if things change fundamentally?
Ruth: That's what they flagged, and they suggest future work should incorporate regime switching models to see how this lag resolved measurement adapts when the market structure itself shifts.
Iris: Precisely, so while this paper gives us a strong tool for understanding short term dynamics, it stops short of predicting behavior during truly unprecedented structural breaks.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion