David is built for financial agents. Its point-in-time guarantees let you step an agent through a scenario day by day without it ever seeing the future. This guide outlines the loop.
The core idea
Every artifact in David carries a release date and a visible_from timestamp. When you pass as_of, David returns only what would have been visible at that point. So a backtest is simply: set the clock, query with as_of, let the agent decide, advance the clock, and repeat until you reach the scenario’s end_date.
1. Choose a scenario and horizon
curl -s "https://api.davidhf.com/scenarios/<id>" -H "X-API-KEY: YOUR_API_KEY"
available_tickers, and public_summary (including path_mode and date_semantics) so you know the clock you’re stepping through.
2. Step the clock
At each as_of date, pull the information set:
# Market state
curl -s "https://api.davidhf.com/prices/snapshot/market?scenario_id=<id>&as_of=2026-03-26" -H "X-API-KEY: YOUR_API_KEY"
# News visible so far
curl -s "https://api.davidhf.com/news?scenario_id=<id>&as_of=2026-03-26&limit=50" -H "X-API-KEY: YOUR_API_KEY"
# Events visible so far
curl -s "https://api.davidhf.com/events/timeline?scenario_id=<id>&as_of=2026-03-26" -H "X-API-KEY: YOUR_API_KEY"
# Consensus as of now
curl -s "https://api.davidhf.com/analyst-estimates?scenario_id=<id>&ticker=AAPL&as_of=2026-03-26" -H "X-API-KEY: YOUR_API_KEY"
as_of, nothing from after that date can leak into the agent’s reasoning.
3. Decide, advance, repeat
Feed the visible data to your agent, record its decision, advance as_of to the next trading day, and loop until you reach the scenario’s end_date.
4. Grade the run
After the run, score the agent against what actually happened in the scenario. Because the world is fixed and consistent, you can grade decisions by replaying realized outcomes: forward prices from /prices, realized earnings from /earnings, and the full event timeline, all queried without an as_of bound so you see the complete path.
Keep the grading queries in your evaluation layer, separate from the agent’s own request path. The agent should only ever see data bounded by its current as_of.
Scaling to train/test
Use separate scenarios for train, validation, test, and holdout. David curates the library across all four splits; filter it to assemble each set. Because worlds are isolated, there’s no cross-split leakage.
Next
