← All speakers

Sid Pednekar

Sid Pednekar leads data and AI at Arcsight Capital at Millennium, building teams and shared capabilities for investment research. His work spans particle simulations, quantitative development and investment data science. His writing examines what an apparently successful AI system still demands from people: repairing data, checking whether comparisons make sense and deciding which findings deserve investigation.

On this page

From particle simulations to investment research

Sid earned a doctorate in chemical engineering at City College of New York, studying how particle sizes and interparticle forces affect flowing suspensions. In these materials, interactions among particles determine whether a mixture flows easily or jams.

With Jaehun Chun and Jeffrey Morris, he investigated how particle-size distributions affect dense suspensions. Their simulations showed how mixtures of different-sized particles can flow more easily than uniformly sized particles at the same solids concentration: the different sizes allow denser packing before the material jams. They also found that mixtures containing just two particle sizes could reproduce important flow properties of more complicated distributions when selected statistical properties matched. The work addressed which details a simpler model must preserve to explain a material’s behavior.

Sid worked at Pacific Northwest National Laboratory in 2018 before moving into finance at Millennium. His early investment work involved quantitative development directly with portfolio managers in Millennium’s central organization. He subsequently led data science for an equities relative-value investment team, working across fundamental research, quantitative analysis and risk management. At Arcsight, his stated focus is scaling decision intelligence and building the data and AI capabilities and teams to support it.

What an AI system really costs

Sid distinguishes a working pilot from a system that an organization can sustain. He also acknowledges the temptation to pursue too many AI opportunities: experimentation can uncover useful ideas, but a proliferation of pilots can consume resources without improving the investment workflow.

  • Count the work around the model. In his account of alternative-data forecasting, a forecast may appear automated while people still repair data after mergers, match records to companies and correct predictions that miss important context. Those interventions belong in the system’s cost. Some can be automated; others require continuing expert involvement or reveal a workflow that needs redesign before it can scale.
  • Account for the reviewer’s time. Generating more analysis can increase the burden on the people qualified to assess it. In his writing on AI returns, Sid gives the example of a growth report comparing paying customers in one period with registered accounts in another. Both figures can be accurately sourced while the comparison remains misleading. Evaluating the return requires asking whose time the system consumes, as well as what the organization could accomplish with the time it releases.
  • Make corrections useful to the next person. Sid advocates shared research infrastructure that retains dataset definitions, known gaps and lessons from successful and unsuccessful uses. A new researcher can then assess whether a dataset suits a question without repeating the investigation that exposed its limits. Corrections should feed back into the work through reusable checks, better context or changed processes, with the supporting evidence preserved.

His AI Engineer talk makes the workflow choice concrete through two hypothetical asset managers. One uses AI to summarize earnings calls and send recommendations to a portfolio manager faster, accepting the risk that summarization discards relevant context. The other starts by asking why material signals arrive too late. Its analysts define investment theses, automated systems continuously gather evidence that supports or challenges them, and analysts investigate consequential exceptions. This assigns people the work of examining evidence that could change a recommendation, rather than simply accelerating every step of the existing process.

1 conference talk

Loading more talks and highlights…