Vaniras
AI/ML stock intelligence platform
A live research product that turns market and company data into probabilistic price forecasts and 0–100 factor scores for 1,000+ U.S. stocks. I train the forecasting models myself and built everything around them: data pipelines, APIs and the web app.
- Role
- Creator · sole engineer
- Status
- Live · free and paid plans
- Coverage
- 1,000+ U.S. stocks
- Horizons
- 3, 6, 9 and 12 months
Overview
A single price target hides how uncertain the future is. Vaniras shows a range instead: for every covered stock, my models estimate where the price could land on four future dates, with a lower, a median and an upper estimate, and the product puts those numbers next to company fundamentals and peer-relative scores.
What the product offers
- Model forecasts: independent 3-, 6-, 9- and 12-month estimates with Lower (P10), Median (P50) and Upper (P90) values, each measured from a stated starting price.
- Scorecard: 0–100 readings for Overall, Value, Quality, Growth, Safety and Momentum, plus Cash generation and Balance strength, compared with sector peers.
- Company analysis: financials, valuation and sector context on an open page for every covered company.
- Discovery: screener, side-by-side comparison, heatmap and watchlists across the coverage universe.
- AI research assistant: answers questions using only the data available to the user's plan.
The language model can explain a forecast. It can never create, replace or recalculate one.
How I built it
- Data pipelines that collect market prices, company financials and macro context and turn them into point-in-time features.
- Quantile forecasting models that I design and train myself, run on a schedule for four independent target dates.
- A scoring engine that compares each company with eligible sector peers on a common, dated snapshot.
- Forecast APIs and backend services that publish dated, versioned snapshots to the product.
- The full web product: research pages, interactive charts, screening, accounts, plans and billing.
- An LLM narrative layer with versioned prompts, compact structured inputs, strict JSON outputs, missing-data rules and token budgets.
Model quality and guardrails
- Point-in-time inputs: every feature uses only data available on the forecast date, with documented publication lags.
- Time-ordered, walk-forward validation with an embargo between training and evaluation. No random train/test splits.
- Out-of-sample evaluation with pinball loss, interval coverage and interval width.
- Invalid or crossed estimates fail closed instead of being repaired for display.
- Numbers and narrative stay separate: the language model only interprets estimates the models already produced.
Vaniras forecasts are model estimates, not financial advice.