Research methodology

Define the question

A number labeled AI confidence is not enough. A WickArk probability refers to a specific event, instrument, holding horizon and research policy, using information available at a recorded cutoff.

For example, positive net profit after five trading sessions is a different event from reaching a profit target before a stop within five sessions. One estimate cannot be relabeled as the other. Entry conditions, the time-based exit and included costs also matter. Long and short scenarios require separate evidence; short success is not simply one minus long success.

In the historical probability workflow's barrier-label research, neither barrier reached within the horizon is TIMEOUT, a retained outcome rather than an automatic zero payoff. If a bar reaches both barriers and their order is unknown, it is marked ambiguous, counted and excluded from that supervised training/evaluation sample—not assigned a favorable ordering. Read ambiguity counts and coverage alongside results; historical labels are not actual fills.

Historical testing is not forward evidence

These evidence types answer different questions. A historical period used to choose a strategy, model or threshold is no longer fresh confirmation of that choice.

Backtest: A simulation of a specified strategy on historical data, under stated costs and execution assumptions. A strong result can reflect overfitting or assumptions that would not hold in practice.
Out-of-sample (OOS): Historical evaluation on a period kept separate from the fitting and selection being assessed. It is stronger than evaluating on the fitting data, but it is still historical evidence.
Walk-forward: Repeated chronological fitting and later evaluation windows. The procedure may change between folds only under its defined rules; future observations must not influence earlier decisions.
Forward evidence: A forecast recorded before its outcome becomes observable. Forward shadow research places no orders; paper execution uses simulated brokerage fills; actual execution uses real fills. Keep those records separate.

Probability calibration

When WickArk displays an eligible probability estimate, calibration asks whether forecasts near that probability historically realized near that frequency on the defined evaluation data.

Read the event definition, evaluation dates, observation counts and reliability bins together. Sparse bins or a changed market environment can make the apparent match unreliable. Calibration is assessed on eligible temporal data separate from the data used to fit the calibration step.

Historical calibration does not guarantee the outcome of a single trade, future calibration or profitability. An estimate can be well calibrated without providing a useful edge after costs.

Stress-test the assumptions

Stress testing asks how sensitive a result is to the configured changes below, where the necessary data and workflow support exist. Inspect the tested ranges and unfavorable results—not just a pass badge or the number of scenarios.

Transaction costs: Higher fees, slippage and applicable financing or borrow costs can remove an apparent edge.
Parameter changes: Nearby supported settings help distinguish a broad stable result from an isolated historical optimum.
Time periods: Changing the evaluation window tests whether one favorable episode dominates the result.
Market regimes: Supported market-state breakdowns can reveal weaker conditions. Small or missing subgroups remain inconclusive.
Execution assumptions: Delays, gaps and available liquidity can change entry, exit and fill outcomes. A simulated fill is not an actual fill.

Model agreement is not correctness

Agreement summarizes disagreement among eligible, compatible forecasts for the same event, horizon and policy. Similar estimates can still share the same blind spot or source of error.

A HIGH, MODERATE or LOW agreement label is not a probability that the trade will win. Outputs with different meanings—such as ranks, return quantiles and event probabilities—cannot be combined as though they were interchangeable. A single eligible forecast does not establish multi-model agreement.

No qualified setup is a useful answer

WickArk can return NO QUALIFIED TRADE SETUP when the configured research requirements are not met. Reasons may include insufficient observations, unavailable calibration, stale inputs, excessive disagreement, weak stress evidence or unsupported costs and liquidity.

Not every stock or policy has an eligible public forecast. The evidence view identifies unmet prerequisites rather than inventing a trade or lowering thresholds to populate a card. Research engine availability alone is not serving approval.

Also ask how selective the process is: how many instruments were considered and how many qualified? A result without its sample size and coverage leaves out important context.

A research target, not a performance claim

The North Star of 70% net win rate and 2:1 average net winner-to-loser payoff is an internal, long-term research target. It is not a claim of historical or forward performance, model accuracy, or achieved customer returns.

The payoff ratio compares the average net winning outcome with the absolute average net losing outcome. A target placed twice as far from entry as a stop is only nominal barrier geometry; it does not demonstrate that realized payoff. Costs, gaps, timeouts and actual execution affect the result.

Research integrity: keep the record

Versioned datasets, strategies, policies and experiment definitions preserve the assumptions behind a result. Issued forecasts retain their original probability, timestamp, outcome definition and horizon; later outcomes are recorded separately, not used to rewrite the original prediction.

Research history retains failed, rejected and unsuccessful work rather than presenting only the winner of a search. Losing issued forecasts belong in the record. Corrections create linked records or versions; they are not permission to silently improve past results. Customer-selected trades are not the complete model-issued sample.

Inspect the evidence, not proprietary internals

WickArk research may use multiple statistical and machine-learning model families. Customer scenario summaries explain the defined event, eligible probability, agreement, validation, robustness and risks without publishing proprietary model identities, weights or features.

Private Research Core recipes, training data and model-selection logic stay behind server-side authorization and explicit public-output boundaries. This page does not fetch private research or hidden holdout results. An explanation cannot create evidence that the underlying artifacts do not contain.

What the evidence cannot promise

Past performance does not guarantee future results. Even carefully separated evaluation leaves uncertainty.

Data limitations: Delayed, missing, revised or incomplete data and limited historical coverage can change what is supportable. Information must have been available at the decision time—not merely present in a later dataset.
Execution differences: Research entry and exit prices are not guaranteed fills. Stops can execute beyond their trigger during gaps or fast markets. Borrow availability and costs are not assured.
Model drift and search bias: Relationships can change. Repeatedly trying ideas can produce impressive historical results by chance. Neither model agreement nor the absence of a warning proves an edge persists.
Short-selling risk: Short losses can exceed the original investment and may be theoretically unlimited. A forecast of a price decline does not establish that a short trade is feasible or profitable after costs.

Understand the evidence behind an estimate.