When Mortality Meets Memory: Pricing Risk in the Long Haul
A mechanism-first reading of how long-memory mortality and interest-rate models change the pricing and risk management of catastrophe mortality bonds.
Cognaptus Insights
Industry thoughts, trends, and commentary.
When Small Coins Roar: Rethinking Systemic Risk in Crypto Volatility Forecasting
A state-adaptive volatility forecasting paper shows why crypto risk desks should monitor spillover roles, not just market capitalisation.
When the Market Speaks: A New Dataset That Actually Listens
FinMarBa reframes financial sentiment analysis by replacing human semantic labels with labels derived from actual next-day market reactions.
From Scroll to Structure: Rethinking Academic Reading with TreeReader
TreeReader shows that the useful AI reader may not be a chatbot at all, but a structured interface that turns papers into navigable, verifiable maps.
Merge Without Mayhem: How Orthogonal Deltas Could Revolutionize Model Composition
A practical reading of MDM-OC, a model-composition framework that uses orthogonal task deltas to merge, update, and unmerge fine-tuned models with less interference.
Mind's Eye for Machines: How SimuRA Teaches AI to Think Before Acting
SimuRA shows why useful AI agents may need less blind action and more internal rehearsal before they touch the browser.
Noisy by Nature: Rethinking Financial Time Series Generation with GBM-Inspired Diffusion
A mechanism-first look at why financial diffusion models may need price-proportional noise, not just larger neural networks.
Seeing is Retraining: How VizGenie Turns Visualization into a Self-Improving AI Loop
VizGenie shows why serious AI visualization assistants need validated workflows, domain-aware vision, and retraining loops—not just prettier prompt-to-code demos.
🚀 All Talk, No Stocks? What Reddit Sentiment *Doesn't* Predict
A close reading of why Reddit sentiment models underperform simpler attention signals in meme-stock prediction.
Forgetting by Remembering: A Smarter Path to Machine Unlearning
A mechanism-first reading of Influence Approximation Unlearning, a faster approximate method for deleting training-data influence without expensive Hessian computation.