METHODOLOGY & SCORING

As of May 2026 — frozen at each published stat update

What the engine measures

Market Sentinel compares an asset with a benchmark and classifies where it sits relative to its historical relationship. The output is a benchmark-relative signal state, not a standalone price target.

The engine asks a familiar analyst question: is the asset trading below, near, or above its normal relationship with the benchmark?

Three signal states

Signal state	Condition	What it means
Recovering	Below the benchmark-relative floor	Catch-up versus the reference is favored, graded at the signal window or the model-implied recovery window
Inflection	Inside the neutral band around the benchmark relationship	No directional prediction. The signal does not have enough edge to grade.
Trending	Above the benchmark-relative floor	Near-floor calls should hold support; extended calls should move back toward the floor over the model window

Inflection calls are excluded from outcome statistics. They make no directional prediction and cannot be graded confirmed or failed.

How calls are graded

A call is graded against the criterion implied by its signal state and distance from the benchmark anchor, using only data available at the signal date.

The public headline is the confirmed-or-partial rate: confirmed means a tight match; partial means the relationship was useful but timing or magnitude was imperfect. Older simple up/down accuracy is retained only as a diagnostic because it discards partial confirmations.

Grading rules

Recovering call is confirmed if the asset outperforms its reference at the 12-month signal window or at the model-implied recovery window for stronger signals.
Trending support call is confirmed if the asset holds above its benchmark-relative floor through the expected support window.
Trending reversion call is confirmed if an extended relative position moves back toward the floor in the expected direction.
Partial outcomes are retained when the relationship moved in the useful direction but timing or magnitude was imperfect.
Inflection calls are excluded — they signal uncertainty, not direction.
Low-conviction calls excluded: signal too close to the neutral band to be directional.
Noise-floor outcomes excluded: relative move too small to distinguish from ordinary market noise.

Nature of results

These are historical replay results, not live managed-account results.

Each call is reconstructed from point-in-time archived data: the engine is run as it would have run on that date, using only data available on that date. No future data is used in constructing any historical call. The calls are generated forward-in-time and graded against subsequent market outcomes — they are not fitted to outcomes.

Results on the track record page reflect the engine as deployed at the time of each backtest run. Methodology changes are versioned; prior published stats reflect prior methodology versions.

Inclusion and exclusion rules

The following tickers are excluded from outcome statistics. Exclusion rules are fixed in the methodology and applied uniformly — they are not selected to improve reported results.

Excluded: cross-currency country ETFs

Country ETFs denominated in USD but tracking non-USD markets (INDA, EWJ, EWZ, VGK, FXI, EWT) mix two different return drivers: local equities and the USD/FX rate. The benchmark relationship is not clean enough to grade. These are flagged as out-of-domain, not wrong calls.

Full exclusion criteria and benchmark selection methodology are documented above.

Excluded: benchmark-contaminated tickers

Indian IT stocks (TCS.NS, INFY.NS, WIPRO.NS) are excluded because they are major constituents of their benchmark (ITBEES). The benchmark drifts with the stock, making the relative floor invalid. This is benchmark contamination, not model failure.

Indian conglomerate tickers with known relationship breaks (ADANIENT.NS post-Hindenburg, ITC.NS under regulatory pressure) are excluded after the break event date. The benchmark relationship does not survive the break.

TATAMOTORS.NS excluded: data unavailable from source during audit period.

What is not excluded

Failed calls are included. All confirmed calls are included. Partial outcomes are published, not hidden. All ambiguous calls where the outcome was within the noise floor are excluded by the noise-floor rule above (uniformly applied). No calls are excluded because they failed.

Benchmark pairs

Every analysis is relative to a reference. The reference is fixed per ticker and is the same for all users:

Asset class	Reference	Rationale
US equities	Sector ETF (XLK, XLF, XLY…)	Same-sector benchmark separates stock-specific performance from sector drift
India equities	Sector Bees ETF (NIFTYBEES, BANKBEES, ITBEES)	Local-market, local-currency benchmark
Crypto	USD (BTC/USD, ETH/USD)	USD as settlement reference
Gold/Silver	GLD/SPY or SLV/GLD	Commodity vs equity or precious metals relative trend
Fixed income	Duration-matched benchmark (TLT/IEF, HYG/LQD)	Duration or credit differential separates rate and credit effects
Currency	Cross-rate pair (EUR/JPY, GBP/EUR…)	Relative currency relationship, not USD-only

Methodological foundation

The engine starts with a fixed historical relationship between an asset and a benchmark. It then measures how far today's relationship is from that anchor and how reliable that distance is.

Uncertainty is highest near the neutral zone and lower when the asset is clearly extended or clearly depressed versus its benchmark. The implementation uses fixed decision rules and point-in-time replay. No rule is tuned after seeing the outcome.

Public validation record

The public product reports the finance audit separately from broader research deposits. The cross-domain validation record is timestamped externally so readers can distinguish the public prior-art record from the commercial implementation.

Five-domain validation record — finance, climate, seismic, biological, and plasma/solar examples.
Uncertainty as aliasing — sampling and neutral-zone foundation.
Five structural gaps — boundary and failure-mode framing.

These links are methodological references, not live trading results. Live-market claims should be evaluated from the track record and audit reports on this site.

What this engine does not do

Does not produce standalone absolute price targets. Price bands are benchmark-conditioned translations, with a flat fallback when the benchmark path is unresolved.
Does not account for earnings surprises, management changes, regulatory events, or other discrete corporate events that break the historical relationship.
Does not personalize to user holdings, risk tolerance, or investment objectives.
Does not execute trades, manage accounts, or route to brokers.
Does not guarantee outcomes. All signals are probabilistic classifications under a fixed model.

Limitations

Relationship break risk: A ticker that breaks from its historical floor (regulatory action, fraud, forced restructuring) can continue to show Recovering even when recovery is no longer realistic. The floor validity check catches persistent breaks but not all event risk.
Look-back dependency: Classification depends on the depth of historical context available. Early calls on newly-listed tickers or tickers with thin history carry higher uncertainty.
Reference validity: The prediction is only as valid as the reference pair. If the reference itself undergoes a major relationship break, both the floor and the prediction are invalidated.
Signal-matched grading windows: Recovering signals use the 12-month signal window or a model-implied recovery window. Trending signals are graded as support-hold or mean reversion, not blanket 6-month continuation.

For informational purposes only. Not financial advice. Benchmark-relative signals are probabilistic indicators, not predictions of future performance. · Full call log · Demo · Plans · FAQ