Methodology
Every probability MyYakan publishes comes from the quantitative engine and maps to a fully-defined, resolvable event. This page describes exactly how the V1 engine defines, estimates, settles, and records each forecast. The numbers are produced by models — never invented — and every recorded forecast is judged by these rules.
The resolvable event
No probability is published without a fully-specified event: an asset, a price level, a 72-hour horizon, and an authoritative settlement price. Every forecast resolves at its own horizon regardless of anything else, so outcomes cannot be quietly withdrawn.
Option A — near-money event
The primary event is whether the close finishes at or above a salient round level within about 0.5% of the current price (close ≥ target). Because the level sits near the price, this probability is naturally close to 50% — it is simple to read and clean to calibrate.
Option B — strong-move events (±0.75σ)
Alongside Option A, two strong-move events are recorded from day one: an up-move (close ≥ a level 0.75 standard deviations above the price) and a down-move (close ≤ a level 0.75 standard deviations below it). These sit well below 50% and carry away-from-50% probabilities whose calibration we track honestly in our append-only record.
What changed (day over day)
Each analysis is compared to the asset's previous recorded slot to show how the usual range and the level of uncertainty moved since last time. This is computed only from already-recorded forecasts — we never restate or invent a past number.
What-if (illustrative move probabilities)
Beside the recorded scenarios, we show an illustrative probability of an up-move and a down-move at a few sizes, from the volatility model anchored to the recorded snapshot. These are exploratory: each is a fully-defined resolvable event, but is NOT part of our scored calibration set, and is labelled as such. They are shown as ranges, never false-precision decimals.
Events and their historical impact
When a high-impact scheduled event (e.g. a central-bank decision, inflation or employment release) falls within an analysis's horizon, we flag it as context — never a prediction of its outcome, and we re-evaluate the read around it. We also show how the asset historically moved around past events of the same type (average move, how often it rose, and the sample size), flagged as merely indicative when the history is thin. Event dates come only from a curated, verified calendar; we never publish an unverified date.
Unscheduled activity
Beyond scheduled events, the engine notices unscheduled unusual activity. When an asset's latest daily move is unusually large relative to its own recent volatility, we show how it resolved after comparable past moves in the same direction — how often it continued, and the typical follow-through over the next few days. This comes only from price and realised history: never an invented cause or headline. It is shown only when the move is genuinely unusual and enough comparable history exists, and thin evidence is labelled as indicative.
Settlement
Every forecast settles on an authoritative daily close for its market, recorded with the resolution. Crypto settles on the exchange daily close — the deepest-liquidity venue, a public number anyone can verify (a median across three or more venues is the planned upgrade). Other classes settle on the reference daily close for the asset, with a stronger per-class source (multi-venue median or official settlement) as the planned upgrade. A class is recorded only once its settlement source is wired — we never record a forecast we cannot resolve — and the published source always names exactly what resolved each forecast.
Where the probability comes from
V1 combines two estimators: a volatility model (EWMA with fat-tailed Student-t innovations) and an empirical model (the recency-weighted historical distribution of past horizon returns). The drift term is class-dependent — zero for the oscillating launch classes (crypto, forex), as their data supports — detailed under Multi-asset calibration. The options-implied estimator (E1) is deliberately deferred and will be added transparently as a labelled new methodology version; its absence honestly lowers source agreement and therefore confidence.
Multi-asset calibration
The same engine serves five asset classes — crypto, currencies (forex), commodities, indices, and equities — and its calibration is validated out-of-sample on independent history for each. Two class-specific corrections, each validated before use: the horizon volatility scales by the number of trading sessions to settlement, so session-traded assets (indices, equities) are not over-scaled; and a small realised-mean drift is added only for persistently-trending classes (commodities, indices, equities), while the oscillating launch classes (crypto, forex) keep a zero-drift model. Any approved asset is analysable on demand; a class is recorded only once its settlement source is wired.
Confidence
Confidence is a separate quality axis — how much we trust the estimate itself — not certainty about direction. It is shown as a band (High / Medium–high / Medium / Low), computed from data quality, source agreement, market-regime familiarity, and known market conditions. High confidence on a balanced probability means a high-quality estimate, not a directional call.
Integrity of the record
Every forecast is written once to an append-only registry as a tamper-evident SHA-256 hash chain — editing or deleting any past record breaks the chain for every record after it. The registry is self-auditing: any edit to a past record is detectable by re-verifying the chain.
Limits — when not to over-trust the numbers
These are probabilities, not predictions: a 20% event still happens one time in five. V1 makes no claim to beat the market baseline and shows no performance numbers until real resolved forecasts exist. The options estimator, a multi-venue settlement median, and event-scenario calibration are deferred and disclosed, not hidden.
Methodology version: v1. Any change to these rules starts a transparent, dated new version — the past record is never silently rewritten.