Model Signals and Accountability

A model signal is a structured read generated from market features. It may say bullish, bearish, mixed, watchlist, or no edge. It may also have a horizon, such as 1-day or 5-day, because a signal that works over one day may not work over five days.

Accountability means the signal does not disappear after it is published. The original timestamp, features, direction, horizon, and confidence should be stored. After the horizon passes, the system should evaluate what happened. Did price move in the expected direction? Was the move large enough to matter? Did the model abstain correctly? Did feature coverage support the call?

A concrete example: the model says CL1 is bullish on a 5-day horizon. That is useful, but not enough. Users should also see whether the 5-day CL1 model has been reliable recently, whether the current features are complete, and whether external evidence like spreads, inventories, and flows support the signal.

The same symbol can also have different reads across horizons. A 1-day model may be reacting to price momentum or a tactical headline. A 5-day model may care more about whether the evidence has time to confirm. That is why the horizon must be shown beside the signal.

A signal also needs a sample context. A model can look strong over all-time history but weak over the most recent 30 resolved signals. Or it can look weak overall but recently improve after a feature set changes. Accountability should show enough context to prevent cherry-picking.

Traders care because a model without accountability is just another black box. A model can be directionally bullish and still deserve low trust if recent outcomes are poor, feature coverage is thin, or the market regime has changed. A model can also have a modest confidence score but deserve attention if the evidence is broad and reliability is improving.

Accountability prevents hindsight storytelling. It shows what the model said before the outcome was known. That matters because market commentary often sounds obvious after the move. A timestamped signal forces the system to live with its prior call.

Looking at the raw model output alone can be misleading. A bullish signal is not the same thing as a buy instruction. It needs horizon, reliability, confirmation, feature coverage, and risk context.

Accountability also changes how users interpret confidence. A confidence number without calibration can be dangerous. A lower-confidence signal with strong recent reliability may deserve attention, while a higher-confidence signal with weak recent outcomes may deserve caution.

A useful accountability view should also separate active and resolved predictions. Active predictions are still being evaluated. Resolved predictions have reached the horizon. Mixing them together can make hit rates look better or worse than they really are.

This is why a useful model page should show both the current thesis and the model's earned trust. The current thesis says what the model sees. The audit trail says whether users should lean in, monitor, or discount it.

Bullish signal with healthy feature coverage: the model has enough current inputs to make a meaningful read.

Recent hit/miss behavior is stable: the model has not been failing repeatedly in the current regime.

External confirmation agrees: price, inventories, spreads, flows, or equities support the bullish direction.

The horizon matches the setup: a 5-day read may be more appropriate when confirmation needs time, while a 1-day read may be more tactical.

Reliability adjustment supports trust: the signal quality remains high after recent performance is considered.

Bearish signal with enough inputs: the model has adequate current feature coverage.

Recent bearish outcomes are credible: similar signals have recently worked or avoided major misses.

External data confirms looseness: inventories build, spreads weaken, flows normalize, or demand evidence softens.

The read survives horizon checks: the 1-day and 5-day views do not violently contradict each other without explanation.

Reliability-adjusted quality remains acceptable: the system has earned enough trust to lean directional.

Signals become mixed when horizons disagree, feature coverage is thin, or external data contradicts the model. A 1-day bullish read and a 5-day no-edge read can both be valid if the immediate price move lacks broader confirmation.

No Edge is especially important in model accountability. If the system has not earned trust recently, or if the evidence is too noisy, the model should be able to abstain instead of forcing a clean label.

Mixed model evidence can also happen when the model is right for the wrong reason. A price move may hit the direction, but the supporting evidence may not match the stated thesis. Accountability should care about that distinction over time.

Neutral model evidence can be valuable when a market regime changes. If the model starts abstaining more often because features are unstable or recent outcomes are weak, that may be a better behavior than continuing to force directional labels.

Accountability should also explain misses. A miss caused by a sudden headline shock is different from a miss caused by weak feature coverage or a stale input. Users learn more when the system names the likely failure mode.

That is why the audit should be visible before users rely on the next signal. A transparent miss can improve trust more than a hidden win, because it shows the system is measuring itself instead of only marketing itself.

Enerlytics tracks raw signal, horizon, feature coverage, confidence, reliability-adjusted quality, recent hit/miss behavior, previous prediction audit history, and current drivers. The goal is to show not only what the model says, but why it should or should not be trusted right now.

The audit history is the trust layer. It should show hits, misses, pending outcomes, No Edge decisions, and reliability changes without cherry-picking only good examples. Historical performance does not guarantee future results, but hiding prior outcomes makes the signal much less useful.