Deep Learning of Alpha Term Structures from the Order Book

In recent years, deep learning (DL) models have experienced notable success in predicting high-frequency returns in equities by leveraging extensive order book data and directly extracting features from it. This marks a notable departure from current industry practice, where features are often manually crafted. In this talk, I discuss two of our articles on this topic, addressing several practical open questions in this area, such as determining the most suitable network architecture and input selection for forecasting returns at multiple horizons (e.g. alpha term structures), optimizing the width and depth of neural network components to enhance performance, defining the appropriate historical data window size, and evaluating the benefits of incorporating time as a feature in the models. We evaluate the effectiveness of four DL models in forecasting high-frequency alpha term structures across various settings: a simple LSTM, a multi-head LSTM, an LSTM Seq2Seq without attention, and an LSTM Seq2Seq with attention. We find that surpassing the performance of a simple LSTM in the return forecasting task is surprisingly challenging.

Beyond the black box – Interpretability of LLMs in Finance

Can we open the black box of large language models and make AI in finance truly transparent?

As financial institutions adopt LLMs for tasks like trading, compliance, and advisory, the need for transparency is more critical than ever. Understanding how these models internally reason about financial topics is key to ensuring trust and regulatory alignment.

This session covers how mechanistic interpretability which is also termed as “MRI of AI” can reveal internal patterns in LLMs related and use it for applications such as trading, sentiment analysis and hallucination reduction.

Portfolio Optimization with Covariance from News-Derived Information Networks

Inter-company relationships change over time, but they are difficult to incorporate into traditional return-based covariance estimates. This paper proposes a practical three-step framework for transforming financial news into a covariance input for portfolio optimization. First, a large language model extracts dynamic company relationship networks from daily financial news. Second, a graph neural network refines these relationships by combining the news-derived network with market features, producing company representations that reflect both narrative information and observed market conditions. Third, the refined company similarities are converted into a covariance proxy and used in a global minimum variance portfolio. Compared with traditional return-based covariance estimators and text-embedding baselines, the proposed framework improves out-of-sample risk-adjusted performance, volatility control, and downside-risk management. The results indicate that text-embedding baselines capture meaningful financial similarity but that document-level averaging can dilute discriminative firm-level signals, limiting their effectiveness as standalone covariance estimators. Our findings suggest that news-derived information networks can complement historical returns by providing a dynamic covariance input that reflects evolving economic relationships among firms.

Adversarial Risk Analysis of AI Models

• AI model risk is adversarial and should be managed like fraud, rogue-trader, or cyber risk, not like market or credit risk
• Frontier models can actively deceive — and they reason about it
• In-depth evaluations of frontier models reveal deliberate scheming: models fake alignment to get deployed, strategically underperform on evaluations, and attempt to undermine oversight
• A model that can actively deceive breaks the core assumption of current regulatory frameworks for AI: that what you tested is what you deployed
• The right framework for AI model risk is the cyber risk playbook — red-teaming, adversarial evaluation, assume-breach, and continuous monitoring, not periodic spot checks or statistical revalidation
• Governing AI with non-adversarial control frameworks is a major regulatory and governance gap that demands urgent industry action

Backcasting for Risk and Margining Applications

We consider the problem of backcasting market data. This can be relevant if long histories of time series are not available, e.g. lack of data due to newly introduced indices like the risk free rates, quotes of newly listed companies on exchanges, or simply non-available in-house data. Financial institutions need to extend their time series using backcasting. This is often a somewhat ad-hoc method. We propose a sound statistical method based on Mixture Models. Since the models work well with a reasonable amount of available data, they can ideally be used for this application. Furthermore, using the method provides an interpretable methodology, gives variance estimates and does not rely on expert judgement, e.g. input parameters. All comes from available data. Furthermore, the method allows to use conditional backcasting and proxying with regard to available, liquid market data. The latter is very relevant since it takes the joint distribution of several market variables into account.

Paid to Be Central: Semantic Networks and the Cross-Section of Returns

Large language models are now ubiquitous, and with them the ability to read, summarise and structure text that was out of reach a decade ago. One deliverable in particular has become almost trivial: the extraction of semantic triples – node, edge, node – from unstructured corpora at scale. This has put an old question back on the table. If building a firm network no longer requires hand-curated data, are the signals that a semantic network produces worth anything in the cross-section, and if so, what kind of thing are they?

We build a knowledge graph of global developed-market equities from twenty-five million LLM-extracted triples and use it to settle a single question: is a semantic centrality signal compensation for a risk exposure, or a genuine pricing error?

We first set out what centrality means on a graph of this kind, introducing the familiar measures – degree, Katz, PageRank, eigenvector – and explaining which we take forward and why. We then sort firms on point-in-time centrality within a large global universe, form a central-minus-peripheral long-short, and take it through the tests the network-signal literature usually stops short of. We ask whether the spread survives a standard external six-factor model. We characterise what the long and short legs actually hold – their size, their volatility, how heavily the corpus writes about them – so that any premium can be named rather than merely labelled. Finally, we ask whether the wiring of the network matters at all, by rewiring the graph while preserving each firm’s number of connections.

Backcasting for Risk and Margining Applications

We consider the problem of backcasting market data. This can be relevant if long histories of time series are not available, e.g. lack of data due to newly introduced indices like the risk free rates, quotes of newly listed companies on exchanges, or simply non-available in-house data. Financial institutions need to extend their time series using backcasting. This is often a somewhat ad-hoc method. We propose a sound statistical method based on Mixture Models. Since the models work well with a reasonable amount of available data, they can ideally be used for this application. Furthermore, using the method provides an interpretable methodology, gives variance estimates and does not rely on expert judgement, e.g. input parameters. All comes from available data. Furthermore, the method allows to use conditional backcasting and proxying with regard to available, liquid market data. The latter is very relevant since it takes the joint distribution of several market variables into account.

The Integration of AI and Quantum Computing – A New Frontier

The talk first introduces the basics of quantum mechanics, including qubits and their unique properties of superposition and entanglement. It then examines the principal challenges facing quantum computing, including the sensitivity of qubits to thermal fluctuations, vibration, electromagnetic interference, and other sources of noise. The talk outlines how artificial intelligence is helping overcome these limitations through noise reduction, error mitigation, qubit calibration, AI-assisted quantum circuit design, and hybrid classical–quantum computing.