How AI Is Changing Investing— with David Trainer

What It Covers

David Trainer, CEO of New Constructs, explains how his firm built a verified fundamental investing dataset over 20 years, partnered with Google Cloud to create a domain-specific AI agent called FinSite, and why reliable data inputs — not general large language models — determine whether AI produces trustworthy stock analysis.

Key Questions Answered

• •Walled Garden AI Architecture: General-purpose AI tools like Claude or Gemini produce unreliable stock picks because they draw from unverified internet data. Domain-specific agents that restrict inputs to audited, validated datasets produce deterministic, trustworthy outputs. Investors evaluating AI tools should ask explicitly whether the underlying data is sourced directly from filings or scraped broadly from the web.
• •Data Integrity Standard: A dataset that is 99% accurate but lacks identification of which 1% is flawed is functionally unreliable for financial decisions. New Constructs validates every data point against original SEC filings with full audit trails — a standard confirmed when Harvard Business School reviewed 350 companies and found zero errors across all adjustments.
• •Core Earnings Edge Signal: New Constructs calculates a "core earnings edge" metric — core earnings minus net income, divided by total assets — to rank stocks by earnings quality. A Bloomberg index built on this signal returned 27% in the most recent year versus 18% for the S&P 500, and a "very attractive stocks" index has outperformed the S&P by 30–40 percentage points over five years.
• •Reverse DCF for Valuation: New Constructs uses reverse discounted cash flow modeling to determine what future profit growth a stock's current price implies. When NVIDIA's price implied a permanent 50% profit decline, the firm flagged it as a long idea. Investors can apply this framework by asking: what does the current stock price require the business to do, and is that realistic?
• •Agentic AI Workflow: Reliable AI at scale requires stacking multiple narrow, domain-verified agents rather than relying on one general model. Each agent must be built on a 100% auditable dataset for its specific domain. New Constructs is scaling its US fundamental dataset globally in 2025, aiming to extend the same earnings-quality and valuation signals to international stocks.