// Research notebook

Scribbles

The messier part of how research actually works — passages I've highlighted, threads I've followed, and bets I've made.

Passages from papers I've been reading. Highlighted, underlined, argued with. Not summaries — just the lines that made me stop.

739× speedup. same accuracy. this is not a small thing.
// Paper StreamTGN · Zhang et al. arXiv:2603.21090

In an L-layer TGN, a new edge affects only nodes within L hops of the endpoints — typically less than 0.2% on million-node graphs. StreamTGN maintains persistent GPU-resident node memory and uses dirty-flag propagation to identify the affected set, reducing per-batch complexity from O(|V|) to O(|A|) with zero accuracy loss. Even a modest reduction in per-query inference latency has massive cumulative impact: (10ms − 5ms) × 10⁸ queries/day ≈ 10⁶ seconds/day.

→ the training papers were optimising the 4% of compute time. inference is the other 96%. nobody touched it until now.
The affected-set insight is obvious in hindsight. A new edge between u and v can only change embeddings within L hops. But every existing TGN system was recomputing all |V| nodes anyway. The reason: the systems were built for training, where you process the full graph, and nobody rebuilt them for inference. 739× speedup was sitting there unclaimed.
// Paper Bottleneck Analysis of Dynamic GNN Inference on CPU and GPU arXiv:2210.03900

This temporal data dependency accounts for the low GPU utilization of TGAT (5%–6%) during the inference process. In JODIE, since the output of the previous time RNNs is to be used as the input to the RNNs for the next time step, RNNs in the following steps need to be stalled until all RNNs in the previous steps finish the computation. The data dependency between RNNs makes the GPU utilization of JODIE also very low (around 1.5%–2.5%). Moreover, the profiling experiment reveals that running inferences on GPU does not outperform the CPU in all batch sizes for DyRep and LDG.

← GPU doesn't even beat CPU for some of these models. that's how badly the compute pattern mismatch is.
5% GPU utilization. You're paying for 100 units of compute and using 5. The reason is sequential batch dependency — batch i+1 can't start until batch i's memory update finishes. GPUs need parallelism to work. Temporal GNNs need sequential consistency. These two requirements are in direct conflict and the literature mostly pretends otherwise.
// Paper Yang et al. · Towards Ideal TGNNs: 10,000 GPU Hours arXiv:2412.20256

Large-scale benchmarking (10,000 GPU hours) has revealed key design principles for ideal TGNNs: the most-recent neighbor sampler and attention aggregator outperform uniform sampling and MLP-Mixer on most datasets. Many of the existing works on TGNN modeling inadequately explore the design space, leading to suboptimal designs. Viewing TGNN models through a performance-focused lens often obstructs a deeper understanding of the advantages and disadvantages of each technique.

→ 10,000 GPU hours to find out that using the most recent neighbors is better than random ones. the benchmark culture discourages this kind of work.
The most-recent sampling finding matters beyond accuracy — it's also more compatible with incremental inference. If you're caching embeddings and only recomputing dirty nodes, most-recent sampling naturally focuses computation on recently active nodes, which are exactly the ones in the dirty set. The design space search and the systems optimisation are converging on the same architecture independently.
compliance cost of the compliance AI — nobody models this
// Analysis Convergences · The Compliance Bill for AI in Asset Management convergences.substack.com · Mar 2026

Building a compliant AI governance program is not a technology project. It is an organizational transformation. The most important thing to understand about AI regulation in financial services is that it is simultaneously more complex and more imminent than most firms appreciate. The EU AI Act classifies AI systems used for credit scoring as "high risk" and introduces additional safeguards — with full application expected by August 2026.

→ "not a technology project" is doing a lot of work here. most firms are treating it as one.
EU AI Act, DORA, GDPR, and SR 11-7 model risk guidance all apply simultaneously to a large asset manager deploying LLMs in compliance workflows. None were written with each other in mind. The compliance cost of the compliance AI is real and almost nobody is accounting for it in ROI calculations.
// Report Capco · AI Transforming Payments & Financial Crime Monitoring capco.com · Feb 2026

Most institutions still run separate stacks and teams for fraud, AML and sanctions. Fraud and sanctions screening typically happen pre-transaction while AML often triggers post-event. Tools remain predominantly rule-based: banks must pre-define scenarios, which leads to poor detection of emerging patterns and a constant trade-off between missing risk (false negatives) and overwhelming operations teams (false positives).

← three separate systems, three separate teams. criminals operate across all three simultaneously.
Fraud typologies that start as card fraud become AML concerns when funds are moved, then sanctions issues if they hit a designated jurisdiction. A model that only sees one slice of the transaction lifecycle misses this by design. The architecture argument for unification is clear. The organisational argument is what nobody wants to have.
// Report Norton Rose Fulbright · AML Enforcement on the Rise amlnetwork.org · 2025

Global AML fines hit $10.4 billion in 2024, surpassing previous records set in 2023. Enforcement is projected to exceed $15 billion in 2026. Norton Rose Fulbright's Boon predicts focus on AI misuse in laundering. Boon states: "Regulators will demand explainable AI in AML systems."

→ $15B in fines and the industry is running rules written in 2003. the business case writes itself.
The explainability demand is the one that matters most architecturally. You can build a GNN that detects suspicious patterns with 98% precision on synthetic data — but if you can't tell the regulator why it flagged the transaction in language that maps to a specific rule, it's useless in a real compliance context. Detection is the easy part. Explanation is the product.
// Report McKinsey · State of AI Trust 2026 mckinsey.com · Mar 2026

Only about one-third of organizations report maturity levels of three or higher in strategy, governance, and agentic AI governance. This imbalance suggests that while technical and risk management capabilities are advancing, organizational alignment and oversight structures are struggling to keep pace with the rapid expansion of AI use. Security and risk concerns are the top barrier to scaling agentic AI. Inaccuracy and cybersecurity remain the most frequently cited AI risks as adoption expands.

← deploying models faster than building governance. that ends badly in a regulated industry.
In finance the consequences of miscalibrated compliance AI aren't just operational — they're regulatory. A model that generates a wrong SAR narrative doesn't waste analyst time. It can constitute a filing failure. Institutions are building the car before the brakes.
"pilot purgatory" — best phrase I've read this year
// Report BCG · For Banks, the AI Reckoning Has Arrived bcg.com · 2025

Most banks are stuck in what the industry has come to call "pilot purgatory" — running dozens of isolated experiments that never scale. The conventional wisdom says the only way out is "rip and replace" transformation: tear out the legacy core, rebuild from scratch, accept 18-month procurement cycles and eight-figure budgets. But this narrative is both paralyzing and wrong. Banks that cannot reason and act in real time across the entire customer journey will not merely fall behind — they risk becoming operationally irrelevant.

→ the rip-and-replace story is used to justify inaction. most useful compliance AI doesn't need greenfield infra
Only 1 in 4 banks is actively using AI to gain competitive advantage per BCG's own survey. The rest are running the same three POCs they started in 2022. The bottleneck isn't technology — it's that compliance teams still have to submit IT tickets to get data access and wait months.
// Paper Khanvilkar et al. · Regulatory Graphs and GenAI arXiv:2506.01093

The system constructs dynamic transaction graphs, extracts structural and contextual features, and classifies suspicious behavior using a graph neural network. A retrieval-augmented generation module generates natural language explanations aligned with regulatory clauses for each flagged transaction. Experiments conducted on a simulated stream of financial data show that the proposed method achieves superior results, with 98.2% F1-score, 97.8% precision, and 97.0% recall.

← simulated stream. always simulated.
The RAG-to-regulatory-clause piece is the actually novel bit here — generating an explanation that cites the specific rule being violated rather than just flagging anomalous behavior. That's what a compliance officer actually needs to file a SAR. 98.2% F1 on synthetic data tells you almost nothing about real transaction networks but the architecture direction is right.
// Paper Park · Reflexivity as Prompt arXiv:2606.00061

Standard AI-assisted forecasting treats the market as an exogenous system. Reflexivity theory holds otherwise: prices shape fundamentals, and every forecaster is a participative agent in the loop it analyzes. We evaluate three frontier models — GPT-5, Claude Sonnet 4.6, and Gemini 3 Pro — under four accumulating zero-shot conditions across two historically distinct episodes: the dot-com bubble (1996–2001) and the global financial crisis (2004–2009).

→ Soros in a prompt. of course someone did this
The framing is genuinely interesting — most forecasting papers pretend markets are physics. This one asks what happens when you give the model a theory of its own participation in the thing it's predicting. Whether the improvement is real or an artefact of the LLM recognising the narrative shape of a boom-bust cycle from training data is a separate question. Writing this one up properly.
// Report Flagright · AI and the Future of AML Compliance flagright.com · 2026

AI-powered solutions reduce false positives by 90–95%, automate labor-intensive compliance tasks, and detect sophisticated money laundering patterns in real time. Financial institutions using AI for AML achieve faster detection, lower compliance costs (from $180+ billion annually), and better regulatory outcomes.

← vendor report so grain of salt — but the direction is right
90–95% is a wild number. The question is what the baseline false positive rate was — AML systems are famously terrible so even a 70% reduction would be transformative. Need to find primary source on this. Also: reducing false positives is only half the problem. False negatives are the ones that get you fined.
this one kept me up. contaminated test sets are everywhere.
// Paper Zhang & Zhang · Hedge Fund LLM Review arXiv:2605.05211

LLMs pretrained on internet-scale text corpora have likely been exposed to financial news, analyst reports, and market commentary from periods that overlap with academic test sets. This creates a form of data leakage that is distinct from look-ahead bias and considerably harder to detect. The model may pattern-match on absorbed associations between companies, sectors, and outcomes — associations that constitute implicit knowledge of the test period — without any single piece of future information being directly accessible. Reported directional accuracy figures may be substantially inflated as a result.

The insidious thing is you can't fix this with a train/test split. The contamination happened during pretraining, before you ever touched the dataset. Standard backtesting hygiene doesn't catch it. You'd need to know exactly what was in the pretraining corpus and when — which nobody publishes.
// Paper Cast-R1 · Tao et al. arXiv:2602.13802

In practice, effective time series forecasting rarely follows a single-pass procedure. Instead, experienced practitioners treat forecasting as a sequential decision process. They examine historical patterns and contextual information, identify informative features, select forecasting models, and reason over intermediate results to assess forecast reliability. As new contextual evidence emerges, predictions are often revised, highlighting that high-quality forecasting involves a series of interdependent decisions rather than a one-shot model inference.

→ this is literally just describing what analysts do. why has nobody built this before
Every architecture paper I've read in the last three years optimises the model. This one says the model isn't the bottleneck — the decision process around it is. If that's right, we've been solving the wrong problem.

Papers that led to other papers. Findings that contradict each other. Threads I followed and where they ended up.

// Rabbit hole 003  ·  May 2026
Starting point

Agentic forecasting: the model makes a sequence of decisions rather than a single inference. Learns to call tools, accumulate evidence, revise. Performance improves. The paper treats this as a pure forecasting architecture question.

what the paper doesn't ask: what happens to market dynamics when many agents are running this kind of sequential decision process simultaneously?
which led to a paper I didn't expect to find
Then I found

LLM-based market simulations (ASFM, TwinMarket) as a new branch of heterogeneous-agent modelling. The paper traces the lineage from 2002 double-auction simulations through HFT microstructure work to contemporary LLM agents — and notes that automated agents already altered liquidity and intraday volatility before LLMs arrived.

if Cast-R1-style agents proliferate in production trading: they all run similar agentic loops, call similar tools, update in similar sequences. correlated behaviour. flash-crash risk looks different when the agents aren't just executing on signals but actively revising forecasts in real time.
the second-order problem the architecture papers aren't thinking about
The implication
Agentic forecasting at scale is a systemic risk question, not just a performance question

The 2010 Flash Crash happened partly because correlated automated strategies responded similarly to the same market signal at the same time. Agentic forecasting systems that share architecture, training data, and tool sets will respond similarly to the same regime signals. The diversity of decision-making in markets is a feature, not a bug. Homogenising it through shared LLM infrastructure is a risk that nobody is accounting for in the current architecture literature.

regulators are thinking about explainability. they should also be thinking about correlated failure modes. not the same problem.
Agentic AI improves individual forecasting. Widespread adoption of similar architectures may degrade market stability. Both things can be true.
// Rabbit hole 002  ·  Jun 2026
Starting point

Feeding Soros's reflexivity theory to GPT-5, Claude Sonnet 4.6, and Gemini as accumulating context improves boom-bust forecasting accuracy. The models that "know" they're participative agents in a self-reinforcing system do better than models that treat markets as exogenous.

the improvement could be genuine theoretical insight. or it could be: the LLM recognises the narrative shape of a bubble from training data and reflexivity framing primes it to apply that pattern. can't tell from the paper.
which led me to: what happens when LLMs generate their own hypotheses?
Then I found

LLM agent reads an experiment trace, proposes falsifiable factor hypotheses, maps them to executable recipes. Ridge-combined portfolio trained 2020–2022 achieves 44.5 Sharpe on out-of-sample crypto. The key constraint: auditable, point-in-time factor DSL with deterministic engine.

reflexivity paper: give the model a theory, it forecasts better. hypothesis paper: let the model generate theories, audit them strictly. the second is more honest about the confound.
the connection: both are really about what the model does with structure
The connection
Scaffolding ≠ understanding — but it might not matter

Both papers show LLMs perform better when given explicit theoretical structure. Whether this is genuine understanding or sophisticated pattern-matching on the provided scaffolding is arguably undecidable. What matters more for practitioners: the crypto hypothesis paper enforces strict out-of-sample testing and auditable decisions. The reflexivity paper doesn't. the theory matters the audit trail matters.

the evaluation methodology is what separates useful from impressive. reflexivity framing is interesting. auditable hypothesis testing is deployable.
Giving LLMs theoretical scaffolding helps. Enforcing auditable, point-in-time constraints is what makes it usable.
// Rabbit hole 001  ·  Jun 2026
Starting point

LLMs behave as "stubborn sloths" when given contradictory financial evidence — they anchor to their pretraining beliefs and resist updating. The paper found that models with strong initial biases experienced measurably higher cognitive conflict when challenged, but still failed to update correctly.

started reading this because of the confirmation bias angle. expected: LLMs are biased toward their training data. got: they're also biased toward whatever position they committed to first in a conversation.
which means LLMs may have persistent "views" about specific securities
Then I found

71.5% detection rate for dealer hedging patterns (gamma exposure, stock pinning, 0DTE flows) using obfuscated prompts. 91.2% of detected patterns materialised in forward returns. The paper argues this is genuine causal understanding, not pattern matching.

wait. paper 1 says LLMs can't update beliefs correctly when presented with contradictory evidence. paper 2 says LLMs are detecting latent market mechanics that materialise in forward returns. these two can't both be fully right.
which directly contradicts the stubborn sloth finding — or does it?
The contradiction
Pattern detection ≠ belief updating

An LLM can be excellent at detecting structural patterns in market data (gamma exposure, order flow imbalance) while simultaneously being terrible at updating a prior view when contradictory fundamental evidence is presented. These are different cognitive tasks. The bias paper tests belief revision. The gamma paper tests pattern recognition.

this matters for deployment. don't use an LLM as an analyst who updates views. use it as a pattern detector on structured data. the failure mode is when you conflate the two.
Use LLMs for structural pattern detection. Don't ask them to revise a view. The tasks look similar. They're not.

Where I think the research is going. Timestamped, confidence-scored, revisitable. If I'm wrong I'd like to know why.

Prediction #005 Made: Jun 2026
Confidence
55%
Within 3 years
A correlated failure event involving multiple institutions running similar LLM-based trading or risk systems will be investigated by at least one G10 financial regulator. It won't cause a crisis but it will accelerate AI governance requirements.

HFT-related microstructure events (2010 Flash Crash, 2012 Knight Capital) triggered regulatory action. The homogenisation of AI architecture in finance — shared LLM providers, similar fine-tuning approaches, overlapping training data — creates correlated behaviour risk. Not identical to HFT correlation risk but structurally similar. 55% because the timeline is uncertain and the event may be small enough to go unreported publicly.

lower confidence on this one. it's the right risk to watch but hard to time. could be 2026, could be 2030.
Prediction #004 Made: May 2026
Confidence
72%
Within 2 years
Most AI regulation research in finance will shift from explainability to auditability. The question stops being "can you explain why it decided X" and starts being "can you prove what it knew when it decided X."

Explainability research assumes the model's internal representations are inspectable and meaningful. Auditability research assumes the relevant question is about the data and context available at decision time — did the model have access to information it shouldn't have had, or not enough information to make the decision responsibly. The EU AI Act's requirements, and the Norton Rose enforcement trend, are pushing toward the latter framing. It's also a more tractable problem technically.

the terminology shift is already starting. "model cards" → "audit trails." same pressure, different framing.
Prediction #003 Made: Apr 2026
Confidence
82%
Within 24 months
The dominant LLM-finance benchmark will shift from directional accuracy to Sharpe-after-costs. At least one major financial ML conference will make this a submission requirement.

The Zhang & Zhang 2026 paper is one of several now explicitly naming the metric mismatch. BCG's deployment critique, and the general maturation of practitioners reading this literature, are creating pressure. Conference programme committees at ICAIF and similar venues are starting to ask for real evaluation criteria. The shift from "does it predict direction" to "would it make money after costs" is the same transition quant research went through in the 1990s. LLM finance is about 5 years behind.

82% confident. it's slow to change but the pressure is accumulating from multiple directions now.
Prediction #002 Made: Mar 2026
Confidence
65%
Within 12 months
Agentic forecasting systems (Cast-R1 class) will outperform single-pass Transformers on financial time series benchmarks with at least two published replications on real market data — but will show lower average improvement on equity returns than on macro/fixed income series.

The Cast-R1 result is architecturally compelling but only tested on general time series benchmarks. The sequential decision-making advantage should be larger in macro and fixed income where regime switches are more consequential and data is more interpretable. In equity returns the noise-to-signal ratio is high enough that the additional decision steps may not help much — and may hurt by overfitting the decision policy to historical regimes.

65% because I could easily be wrong about the equity vs macro split. that part is a guess.
Prediction #001 Made: Feb 2026
Confidence
78%
Within 18 months
The first major AML enforcement action citing AI-generated false negatives as a contributing factor will happen before end of 2027.

Institutions are deploying ML-based transaction monitoring faster than they're validating its false negative rates on novel typologies. Regulators are already signalling that AI doesn't excuse non-detection — Norton Rose Fulbright's 2025 enforcement report explicitly flags this. The gap between synthetic-data model performance and live-network performance is real and largely unmeasured internally. One institution will get caught.

the timing is what I'm uncertain about. the event itself feels close to inevitable given current deployment patterns.