Market bars and trend signals

The FinanceMarket method serves the market-signal layer of the finance program (ledger rows EH-411, EH-413 to EH-418, EH-420 and EH-421). It covers typed OHLCV bars over the time-series store, deterministic indicators, a per-series trend signal with its flip records, calibrated flip confidence, and the backtest-run provenance record, server-side chart decimation, and the analysis-snapshot record behind shared analysis links.

Everything here is informational only. No op, record or result authorises an order. The engine contains no language model: explanations of a flip belong to the application layer, which must show the mathematical trigger first.

Ontology (finance-v1)

crates/eg-core/ontology/finance-v1.ttl is a thin core module in the style of the world-model modules. It sits under the BFO layer and maps (never imports) to FIBO and Wikidata. It declares ten market classes:

Class	Placement	Notes
FinancialInstrument	generically dependent continuant	Declares the class energy_geopolitics's Commodity already specialises.
Venue	System	A place of trade.
Listing	generically dependent continuant	One instrument, one venue, one quote instrument and one listing type.
BarSeries	Dataset	Names the store series that holds the bars.
IndicatorSpec	generically dependent continuant	Identified by a version and a parameter hash.
SignalState	generically dependent continuant	The latest state of one signal.
TrendFlip	Event	revisesFlip links a revision to the flip it revises.
MacroEvent	Event	A policy action with its announcement time.
BacktestRun	generically dependent continuant	The provenance record of one backtest.
AnalysisSnapshot	generically dependent continuant	An immutable, content-addressed record of one shared analysis.

A FinancialInstrument may carry one assetClass from a closed vocabulary: crypto, stock, etf, fund, commodity, forex, index or bond. A BarSeries also carries its tickSize (required) and volumeStep (optional, default 1), both xsd:decimal. Bar prices are integer ticks, so a reader needs the tick size to show a price.

It also holds the eight trading classes that used to live in company_infra (folded in EH-517, IRIs unchanged): TradingStrategy and TradingDebate (processes), ExchangeBackend and PortfolioPosition (independent continuants), and the records TradingSignal, BacktestResult, VersionedOrder and RiskSnapshot (generically dependent continuants, like every other record here). A BacktestResult is an external backtester's metric summary; a BacktestRun is the engine's sealed, content-addressed record. VersionedOrder documents an order; nothing here authorises one.

finance-v1.shapes.ttl holds the ABox shapes. Closed vocabularies such as the listing type, the price basis, the data status and the direction are sh:in lists, so a misspelled value fails validation. Timeframes and hashes are checked by pattern. With these two documents the core catalog holds 37 of its 64 sources.

Bars (EH-413)

A bar is one store point. ts is the bar's open time, and the fields follow codec::BAR_FIELDS:

• the close time and the known_at time, each split into 32-bit halves;
• the open, high, low and close, as integer ticks;
• the volume, as integer units;
• the status (provisional or final) and the revision.

Every field is an integer that an f64 holds exactly.

The store only appends, and multiple versions of one bar coexist at one timestamp. A correction is therefore a new point with a higher revision. encode_points writes records in this layout for TsAppend. resolve decodes stored points (and takes plain records) and builds the view as of any time: for each open time it takes the highest revision known at that time.

resolve refuses these inputs:

• two different bodies at one revision;
• a provisional version that will revise a final bar;
• a final bar that claims to be known before it closes;
• overlapping bars.

rollup folds whole bars onto a calendar timeframe. A rolled bar is final only when all its parts are final and the caller's watermark has passed the end of its bucket.

There are two kinds of trading calendar:

• utc24x7 aligns days to UTC midnight, weeks to the ISO Monday, and months to the civil month.
• Exchange calendars are entirely data: offset spans that cover daylight saving, one session, the trading weekdays, holidays and early closes. They read no time-zone database, so they replay exactly.

Indicators (EH-414)

All kernels are incremental: each closed bar is one step. The kernels are integer-only; values are in milli-ticks and every division rounds half away from zero. The kernels are:

• Wilder ATR, seeded by the mean of the first period true ranges;
• the ATR trailing-trend line (SuperTrend), on a raw or Heikin-Ashi basis;
• SMA and EMA (the EMA is seeded by an SMA);
• the 20/21 band;
• Heikin-Ashi candles.

The 200-week SMA is an SMA of period 200 over bars rolled up to weekly.

The trailing line follows the public TradingView ta.supertrend recurrence:

• The upper band only ratchets down and the lower band only ratchets up.
• The first computable bar is bearish, and no flip is recorded there.
• A flip needs the close to be strictly beyond the band. A close equal to the band holds the current direction.

The fixture crates/eg-compute/tests/fixtures/finance_market/reference.py is an independent Python implementation. It writes golden.json, and the Rust golden tests must reproduce every digest bit for bit on every build host.

supertrend_companion.pine, beside it, is the TradingView companion: a Pine v5 indicator that replays the same integer Wilder ATR and trailing-line recurrence (raw or Heikin-Ashi basis) on the chart's own bars, plots TradingView's ta.supertrend next to it, and counts the bars where the two disagree in direction or by more than a tick on the band. Its defaults are the fixture's super_trend_10_3; tests/test_finance_pine_companion.py keeps the defaults, the fixed-point scale and the recurrence rules in step with the reference. It plots only and never places an order.

Signal, flips and the scanner (EH-415)

SignalKey is built from five parts:

• the listing;
• the price basis;
• the timeframe and calendar;
• the indicator version;
• the parameter hash.

signal_advance moves a SignalState forward by final bars that come after its last bar. Each bar is one O(1) step from the state's exact kernel checkpoint. A bar at or before the last bar is refused: that change needs a replay.

signal_replay processes every bar version in the order it became knowable:

• An appended bar advances the signal incrementally.
• A revised or back-filled bar triggers a recomputation, which appends revision records. A changed flip gets a new emitted record that revises the old record. A flip that no longer exists gets a retracted record.

No record is ever edited in place. A flip's event_id is derived from the signal key and the bar close, so replays are idempotent. A replay as_of a past time reproduces exactly what was known at that time.

signal_scan keeps one latest state per key and never lets a state regress. It counts states against a stated denominator of one state per key, and orders rows by the most recent flip.

finance::market::events turns each flip into a finance.flip event for eg-stream's CEP engine. This lets compound signals, such as weekly-then-daily agreement, run on the same NFA as the live standing queries.

Calibrated flip confidence (EH-417)

flip_confidence scores a candidate flip at one horizon, using the smoothed follow-through rate of its feature cell. Calibration uses the Decide layer's own kernels:

• the conformal minimum-sample gate;
• a Clopper-Pearson drift check between the earlier and recent halves of the history;
• a class-conditional binary conformal set built on leave-one-out scores.

Only a singleton set is a claim. In every other case the op abstains with a typed reason:

• data that is not valid (warming up, stale or unavailable);
• history thinner than the gate;
• an unsupported regime;
• drift;
• an ambiguous set.

Backtest-run records (EH-418)

backtest_run seals a draft into an immutable, content-addressed record. Sealing requires complete provenance:

• the signal keys;
• the data revisions;
• a point-in-time universe;
• the costs;
• the fill rule.

Every fill must pass the no-look-ahead check: the signal was known at or before the fill, and the listing was in the universe at the time of the fill.

The engine's existing kernels compute the mandatory validation outputs: purged combinatorial CV, the deflated Sharpe ratio, and the probability of backtest overfitting. The caller never supplies these values.

backtest_run::verify re-derives a record from its own draft. A revised run is a new record whose supersedes names the old digest.

Chart decimation (EH-420)

decimate thins bars, and indicator series aligned to them, to a pixel width of 16 to 8,192 columns. Run it after the indicators, which are computed over the full history. Bars are assigned to columns by open time:

• Each column's bars fold into one candle: the first open, the highest high, the lowest low, the last close and the summed volume. This is M4 for candles, so the values a column draws are exact. A column is final only when every bar in it is final.
• Each line keeps its first, last, lowest and highest point in every column. A band keeps both of its lines.
• A trailing line also keeps every change of direction, with the point before it, so the colour boundaries of the line stay exact.
• A Heikin-Ashi series folds one candle per column, like the bars.

A request that already fits comes back unchanged with decimated: false. The arithmetic is integer-only, so the result is the same on every target.

Analysis snapshots (EH-421)

analysis_snapshot seals an analysis as an immutable, content-addressed record. Sealing refuses the draft unless it is self-consistent:

• The signal key must be the key of the spec over the series.
• Every flip must belong to that key, carry its derived event id, lie inside the bar window and follow the previous flip in time.
• The source revision must be a sha256: digest of the bar versions read.
• Every claim must cite at least one source. A source is titled and located by an http(s) URL or an engine record reference. A claim with no source is refused with UNSOURCED_CLAIM.

The engine stamps three notices on every snapshot: informational only, the hallucination warning, and the mechanical trigger. A caller can neither omit nor reword them, because they are part of the digest. A snapshot never includes positions (excludes_positions is always true).

To check a stored record, seal its draft again and compare the digests. An application stores the record as an AnalysisSnapshot graph node, created only if absent, whose id derives from the digest; analysisOf names the listing and analysisDigest holds the digest. It shares the record through a control lease of kind share.read that names the digest. The lease is bound to the tenant, has an expiry and can be revoked.

Signal models (EH-423)

FinanceSignalModels is a separate method that holds two closed-form models. They moved here from agent-utilities so that each model has one owner. Both are pure compute and informational only.

bayes_fuse fuses directional calls (1 up, -1 down, 0 no call) into P(up):

• A measured prior (directional_accuracy, standalone_sharpe, pbo) seeds a source with weight accuracy × sharpe, clamped to [0, 1].
• A prior is dropped when it is overfit (pbo > max_pbo) or has no edge (sharpe <= min_sharpe).
• A source that calls without a seeded prior takes the default weight and accuracy.
• Each call moves the probability toward P(up | call) by the source's weight. Calls are applied in source-name order, so the answer does not depend on how the caller built its map.

insider_equilibrium models the Kyle insider under a dynamic legal-risk hazard (Qiao & Xia, arXiv:2605.27684). It is a surveillance-design aid, not a trading tool. It returns three parts:

• The equilibrium. The optimal intensity is β* = (Σ − eκC) / (2Σ(λ + eκr)), floored at zero. The answer includes the detection hazard, the expected profit, the penalty and the binding lever.
• The schedule. Enforcement decays with the remaining window, so the insider accelerates toward its end.
• The penalty policy. The comparative statics in the criminal and civil penalties, the criminal cost that suppresses the insider (null when no finite cost does), and a typed verdict: enforcement_gated, criminal_suppresses or criminal_is_the_lever.

The validation outputs use only correctly rounded IEEE operations and the pinned soft-float kernel (eg_numeric::detkernel::math), never the platform libm, powi or powf. A sealed digest is therefore the same on every host, and the test a_sealed_backtest_run_digest_is_pinned_across_hosts pins one.

The deflated Sharpe ratio is a normal tail probability, often a very small one. It comes from detkernel::kernels::normal_cdf, which evaluates the pinned erfc directly and never forms 1 + erf or 1 - CDF, so it keeps full relative accuracy far into the tail (checked against mpmath at z = -3 to -10). The Diebold-Mariano p-value uses normal_sf the same way.