Leveraging Data Insights From Risk Management to Marketing Optimization.

Thanks to everyone who joined our meetup co-hosted with JetBrains last Thursday! It was wonderful to reconnect after our September conference. We hope you enjoyed the talks and had some insightful conversations.

Below is a summary of the talks & key takeaways:

Jesse Koreman and Kally Chung’s Talk: 𝐌𝐏𝐋𝐗 - 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐃𝐫𝐢𝐯𝐞𝐧 𝐑𝐢𝐬𝐤 𝐌𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭 focused on three core issues with the current system:

✦ Timing Mismatch: Inaccurate timing in risk assessments.
✦ Inaccurate Assumptions: Flawed risk assumptions raised concerns about insufficient deposit holdings.
✦ Merchant Dissatisfaction: Inaccurate assessments strained merchant relationships, especially regarding deposit requirements.

Using a data-driven approach with survival analysis and nonparametric curve fitting, Jessie and Kally’s team achieved better accuracy, reduced deposit requirements, and improved merchant satisfaction, demonstrating MPLX’s strong impact.

Bas Stinenbosch's talk 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐞 𝐲𝐨𝐮𝐫 𝐦𝐚𝐫𝐤𝐞𝐭𝐢𝐧𝐠 𝐞𝐱𝐩𝐞𝐧𝐬𝐞𝐬: 𝐀𝐧 𝐨𝐯𝐞𝐫𝐯𝐢𝐞𝐰 𝐨𝐟 𝐝𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭 𝐦𝐞𝐭𝐡𝐨𝐝𝐬 discussed the evolution of marketing mix models, from linear regression to Bayesian models, to optimize spend across channels. Challenges include limited weekly data points and privacy regulations, complicating multi-channel attribution. Bayesian modeling, by incorporating prior knowledge, offers actionable insights with limited data. Bas recommends combining marketing mix modeling, multi-touch attribution, and A/B testing for optimal marketing strategies.

Jodie Burchell, PhD's talk 𝐋𝐢𝐞𝐬, 𝐃𝐚𝐦𝐧𝐞𝐝 𝐋𝐢𝐞𝐬, 𝐚𝐧𝐝 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 explored why hallucinations in LLMs happen and ways to mitigate them. LLMs, powered by vast datasets and transformer architecture, can produce impressive but sometimes misleading content. Hallucinations fall into two categories: faithfulness errors, where the model strays from the source, and factual errors, where it confidently shares inaccurate "knowledge." Solutions include prompt-tuning, self-checking, and grounding responses with live data. While hallucinations may persist, understanding and reducing their risks ensures LLMs are applied responsibly.


𝐁𝐨𝐧𝐮𝐬 𝐟𝐫𝐨𝐦 𝐉𝐨𝐝𝐢𝐞
For those in the small group chat after the talk, you already have these 😉 ! Sharing with the larger community:

✦ Another talk by Jodie: Mirror, Mirror: LLMs and the Illusion of Humanity Watch: https://lnkd.in/eZzVFC5y

✦ On the measure of intelligence
https://lnkd.in/fcWYgwr

✦ Levels of AGI for Operationalizing Progress on the Path to AGI
https://lnkd.in/dBNwvCjZ