Continuous improvement
TECHi Intelligence is not a finished black box. It is a public research system that should get clearer, better sourced, and more useful as the app matures. This roadmap shows what is live now, what we are improving next, and what will stay off-limits.
Methodology reviewed May 28, 2026 · Validation dataset updated Sep 23, 2026 · Model v1.0
Live in v1.0
Field-level source visibility
Quote pages name the provider, the as-of time, and whether a figure is provider-reported or a TECHi formula. Source visibility is no longer future work.
September validation window
The validation dataset was updated September 23, 2026. The public replay scores price direction through six months on stored closes. That window is longer than the May methodology review, and it is still not out of sample.
Same-sample simple baselines
Each Signal page scores buy-and-hold, a random directional call, and 20-session momentum on the same closes as the price replay. Those comparisons are not a published universe result.
Public factor weights and stance bands
The methodology page lists the live weights. Constructive, Neutral, and Cautious are research states, not instructions.
Historical setup factor, separate from validation
The 0–100 setup score is named Historical setup factor. The word validation is reserved for the price-tested replay.
Price-direction replay through six months
Stored closes replay 1-day, 7-day, 15-day, 30-day, 3-month, and 6-month price-derived calls. That replay is on the Signal page.
Calibration gate on horizon percents
A horizon past six months is an illustrative fundamental scenario and shows no percent. Inside six months, a high calibration error shows direction only.
Source notes on the quote page
Signal rows show drivers, formulas, coverage, and the price-test window. Provenance on the quote page names the provider and the as-of time.
What improves next
The S&P 500, a sector ETF, and analyst consensus are still not published. The replay does not contain those series on the same dates. A universe average of the simple baselines is not published either, and neither is a volatility-adjusted result. New factors stay paused until those three baselines have a stored result.
The current replay is not out of sample. A walk-forward pass with a true holdout period waits behind the unpublished market and consensus baselines. It is not a new factor.
Survivorship controls and look-ahead checks are not in the public replay. A later close used as both a feature and an outcome would fail that check. Those checks are not a new factor.
Software, semiconductors, banks, utilities, ETFs, and crypto-linked names still share one yardstick. Sector bands stay paused until the S&P 500, sector ETF, and analyst-consensus baselines have a stored result.
Guardrails
A cleaner interface must never hide weak coverage, stale inputs, or a missing filing.
TECHi will keep improving provider coverage, filing reads, and derived fields, but the page must still tell readers when the best available input is imperfect.
Material changes to weights, bands, validation, or confidence rules should land with a version note instead of disappearing into the codebase.
