AION Veritas — Model Specification

Published 22 June 2026 · AION Analytics (India) · Lokesh Gupta · Public validation record →

This page is the canonical public definition of AION Veritas. It explains what the model is, what inputs it draws from, what it outputs, how reliability is graded, and how calibration is maintained. It does not publish live band values — those are subscriber-only. The public validation record (all historical receipts) is at /veritas/validation.

1. What Veritas is

AION Veritas is a pre-market probabilistic model that produces an opening-window band for NIFTY 50, BANK NIFTY, and SENSEX before each Indian cash session. It is published by 08:55 IST on trading days. A band is a calibrated range: a lower bound, an upper bound, and a midpoint. A directional lean (BULLISH / BEARISH / NEUTRAL) accompanies the band with an explicit reliability grade.

Veritas is a probabilistic information tool. It does not issue trade recommendations, price targets, or stop-loss levels. It answers a narrower question: given everything knowable before the market opens, where is the opening window most likely to land? The honest answer includes a range and a reliability score — not a single number.

2. What Veritas consumes (inputs)

Veritas draws on three classes of pre-market signals assembled before 08:46 IST each trading day:

• Pre-open microstructure: GIFT NIFTY 5-minute candle series from the early morning (08:00–08:40 IST). Gap, volatility, range, trend, and momentum over this window are extracted as structured features.
• Market-state context: India VIX (implied volatility regime), prior session close levels, US overnight session moves. These establish the ambient uncertainty envelope.
• Macro event intelligence: The AION Indian Market Intelligence layer classifies overnight financial headlines into structured macro-event signals — covering RBI/policy, crude oil, rupee, risk sentiment, and liquidity channels — and routes their sector-level impact into the feature assembly. On days with no significant headlines, this block contributes neutral-strength signals, not silence.

Feature assembly completes no later than 08:46 IST. The model runs on the assembled feature set and the published band is available by 08:55 IST. The model does not consume intraday price data after the session opens.

3. Output — what the band is

Each Veritas output consists of two independent components:

Opening-window band (the range track)

The band expresses where Veritas expects the opening window to land, in the form of a probability interval:

• Lower bound — the floor of the expected opening range
• Upper bound — the ceiling of the expected opening range
• Midpoint — the central estimate; on low-conviction days, treat this as context, not as a directional anchor

The opening window is defined as the first 15 minutes of the cash session (09:15–09:30 IST). Whether the band "captured" the actual open is verified same-day and published at /veritas/validation.

Directional lean (the direction track)

A separate directional model produces a three-class lean — BULLISH, BEARISH, or NEUTRAL — with an associated confidence score and an explicit reliability grade:

• STRONG — confidence is above the model's minimum threshold; the lean has discriminative value for the session
• WEAK — confidence is below threshold; the model is explicitly signalling uncertainty. On WEAK days: use the band, ignore the lean.

4. Reliability grading

Veritas explicitly reports whether each session's output should be considered reliable for its two components:

These grades are not cosmetic — they are calibrated against the historical record. Days with higher range confidence have materially higher band-capture rates in the validation receipts. The model uses its own uncertainty output to signal when to trust it less.

5. Regime awareness

Veritas is aware of the prevailing India VIX regime at the time of each prediction. In high-VIX environments, expected opening ranges are wider — the model does not force a narrow band when volatility structure says uncertainty is elevated. The VIX regime affects band width. It does not mechanically determine direction.

The interaction between VIX regime, pre-open GIFT structure, and macro event tone is what produces the band. No single input dominates unconditionally.

6. Calibration and honesty

The band is produced using conformal calibration: a post-training statistical procedure that adjusts band width based on how the model performed on held-out live data. The calibration is not static — it is automatically updated on a trailing window of recent sessions. This means the band width adjusts over time as the model's live behavior is observed.

Design coverage target: 80%. The band is calibrated with the goal that the actual opening lands inside the band in approximately 8 of every 10 sessions. Verified live coverage: ~72% — measured unconditionally across all sessions in the public validation record from February 2026 to present. We publish the gap between design target and verified live performance rather than the target alone. The 8-percentage-point shortfall is in the public record, not masked behind a headline claim.

The model does not selectively publish sessions. Every session — including misses, narrow bands on big-move days, and WEAK-direction sessions — is in the permanent public record at /veritas/validation. This is the trust architecture: not claims about accuracy in the abstract, but a verifiable daily audit trail.

7. Worked examples (illustrative)

These are illustrative examples using approximate values consistent with actual model outputs. They are not live subscriber data and do not represent any specific session.

Example A — Strong conviction session

How to read this: the band is narrow and the range confidence is high. The directional lean is STRONG, meaning the model sees enough structure to support a bullish view. On a day like this, the midpoint is informative. The subscriber sees band + lean + confidence. Whether the band held is verified same-day.

Example B — Low conviction session

How to read this: the band is wider. The direction is flagged WEAK — the model has enough information to publish a band but not enough directional structure to justify a lean. The correct reading: trust the band, ignore the midpoint as a directional signal. The band still captured the open.

What the reliability grades mean in practice

Directional accuracy is published by reliability grade, not as a blended figure. The STRONG/WEAK split is the point: the grade is itself the signal.

A blended accuracy figure across all sessions would understate STRONG-day quality and overstate WEAK-day quality simultaneously — which is why we do not publish one. The reliability grade is the model's own forecast of its directional confidence. It is calibrated to that forecast: STRONG days are right significantly more often, and WEAK days are not claimed to add directional value.

8. Access — subscription

9. What is free — the public validation record

The live forward band requires a subscription. What is permanently free and public:

• Every historical session receipt at /veritas/validation — including band bounds, whether the band captured the actual open, directional lean and its actual outcome
• This model specification page
• The educational note on the probability-band lineage (Bachelier to BSM)
• The Veritas model card

The validation record is the trust layer. It is not a product preview — it is a permanent audit trail. A prospective subscriber can read it, count the misses themselves, and make an informed decision.

10. What Veritas is not

• Not a trade recommendation. The band says where opening prices are likely to land. It does not say to buy, sell, or hold anything.
• Not an API for free real-time data. It is a probabilistic model product behind a subscription gate. There is no free callable endpoint for live bands.
• Not perfect. The validation record is the proof: some sessions miss, some directions are wrong. The model publishes its own uncertainty. A WEAK direction flag is the model telling you it does not have enough signal today.
• Not a backtested simulation. The published validation record is forward-looking from the date of first publication. No historical results have been fitted to look better than they were.