How a forecast is made
Each step is a separate agent with one job. The forecast is stated and time-stamped before its outcome exists.
- CollectCollectors capture prices, news and filings as they are published, each item stamped with its capture time.
- AnalyzeOne analyst agent per source type reads its inputs and writes a sourced assessment.
- ForecastThe forecaster combines the assessments into a forecast with a stated horizon and probability.
- AuditAn audit agent checks that every claim in the rationale traces to an input that existed at forecast time.
- ScoreWhen the outcome is known, it is added to the record and the forecast is scored.
Every forecast is a record
Stored at the moment it is made, so it can be scored and replayed.
- Time
- When the forecast was stated, and the horizon it covers.
- Inputs
- The exact items each agent saw, with their capture times.
- Model version
- The model family and pinned version behind every agent.
- Rationale
- The forecaster's reasoning, with references to its inputs.
- Outcome and score
- Added when the outcome is known. Earlier fields stay as written.
How forecasts are scored
Each forecast is fixed before its outcome exists, and every forecast is scored, misses included.
Out of sample
Forecasts are scored only on periods and events the system never saw while it was built or tuned.
Baselines and random controls
Each forecast is compared with simple baselines, such as no change or the historical rate. The same pipeline also runs on randomly chosen dates, and a result counts only when it separates from those controls.
Calibration
Stated probabilities are checked against observed frequencies, so events called likely happen about as often as the system said.
Multiple-comparison correction
When many hypotheses are tested at once, significance thresholds are corrected for the number of tests, so a lucky result cannot pass as skill.
Event and news response
How prices respond to what people and organizations say in public.
We study price responses to public communications: statements, announcements and filings. The work runs on a time-aligned corpus that pairs each statement with the price moves around it.
Every event window is compared with randomly drawn windows from ordinary periods, so a response counts only when it stands apart from normal movement.
The same corpus supplies the dynamic news feed in our agent simulations.
Cross-venue price discovery
We measure how price moves travel between regulated futures and decentralized perpetual futures.
Lead and lag
Which side moves first and how quickly the other follows, measured in both directions.
Out-of-sample tests
Relationships found in one period are tested on later periods they were not fitted to.
Independent reimplementation
Each result is rebuilt from the raw data by a second, separate implementation before it is reported.
The data underneath
Every forecast traces back to these datasets. Each one states its sources and scope.
- U.S. stocks and futures
- Trade and quote history from 2018, in our research warehouse.
- Decentralized perpetual futures
- Order books rebuilt from snapshots and book diffs, with fills, liquidations and funding.
- Public communications
- Statements aligned in time with price moves, with random-baseline controls.
- GPU and neocloud pricing
- Published prices and availability for 11 GPU classes. The methodology captures them from provider pricing APIs and published pricing pages.
What we build for clients
Forecasting systems set up on your questions and your data, scored the way we score our own.
Market-intelligence agents
Agents that follow the sources you choose, report what changed and cite where each point came from.
Forecasting and backtesting pipelines
Collectors, analyst agents, a forecast log and scoring, built around your questions. Backtests replay only what was known at each point in time.
Event studies
How prices, demand or other measures responded to announcements and filings, compared against random-baseline windows.
GPU demand and capacity forecasting
Forecasts of GPU rental prices, availability and capacity needs, built on the GPU cloud pricing data behind the Rillor Compute Index.
Systems run in your environment or on a private AI cloud we operate for you.
Notices
Rillor is not a registered investment adviser or commodity trading advisor and does not provide investment or trading advice. Research, forecasts, data and software described on this site are for research and engineering use. Nothing here is an offer or recommendation to buy or sell any security, commodity interest or digital asset. Past or simulated results do not indicate future results.