Weekly 3x3: CME to launch Solana futures. Texas needs reactors. ChatGPT beats DeepSeek on stock predictions.
My top reads in markets, tech, and AI research this week.
MARKET MOVES
CME Group plans to launch Solana futures | February 28th 2025
Institutional demand for crypto continues to grow. Reuters
ECB's multi-trillion payments breakdown sends shudders through Europe | February 28th 2025
Breakdown on payday draws attention to the interconnectedness and resilience of the financial payments system. Reuters
Why power prices can go negative, and what it means | February 28th 2025
This article from Bloomberg explains that prices can go negative when the cost of holding or disposing of a good exceeds its value, leading sellers to pay buyers to take it off their hands. Bloomberg
TECH TALK
AI will upend a basic assumption about how companies are organized | February 28th 2025
As AI reshapes workflows, companies may need to rethink traditional hierarchies and organizational structures. Bloomberg
Texas needs equivalent of 30 reactors to meet data center demand | February 28th 2025
Surging data center demand in Texas highlights the growing energy challenge. Bloomberg
Blue Origin's all-women flight | February 28th 2025
Katy Perry adds star power to space mission. Bloomberg
RESEARCH RADAR
ChatGPT and DeepSeek: Can they predict the stock market | February 14th 2025
A study finds that ChatGPT can predict the stock market and macroeconomy from Wall Street Journal data, outperforming other models. Its predictive power comes from investors’ underreaction to positive news. ArXiv
Re-routing in test-time for multimodal mixture of experts | February 27th 2025
Vision encoders often underperform compared to language models. Using a mixture-of-experts approach helps, but the router that mixes expert representations isn't always optimal. This paper proposes re-routing in test-time by aligning routing weights to correctly predicted samples in a neighbourhood of the test sample. ArXiv
Expect the unexpected: Failsafe long context QA for finance | February 10th 2025
Writer introduces FailSafeQA, a benchmark testing the robustness of LLMs in finance. Results show that while some models handle input changes well, they struggle with hallucinations. Writer's Palmyra-Fin-128k-Instruct model is the most compliant but falters in 17% of cases, while OpenAI o3-mini fabricates information in 41% of cases. The benchmark highlights the need for improvement in LLMs for financial applications. ArXiv