Changelog

These notes describe fintech-algorithms@0.13.0, the version this site was built from. Anything that can break a caller is called out under the release that introduced it.

All notable changes to fintech-algorithms are recorded here.

The package is pre-1.0. Breaking changes go in the minor version, additive and corrective changes in the patch — so 0.11.0 → 0.12.0 can break you and 0.12.0 → 0.12.1 cannot.

0.13.0 — 2026-08-17#

The catalog and the package finally hold the same set. 675 topics across 17 domains, with nothing skipped and nothing held back — the first release where every algorithm that exists is also one you can install. Nothing is removed: every subpath published in 0.12.0 still resolves.

Technical Indicators is the headline for most people: it shipped 37 of its 137 topics in 0.12.0 and now ships all 137. ALMA, T3, VIDYA, FRAMA, TRIX, Klinger, the Hilbert and Ehlers cycle tools and the rest are installable for the first time.

One thing that can break you#

matching-engines-and-venue-logic/continuous-matching/hybrid-pro-rata-time-matching was publishing its family's calculate(topicId, inputs) dispatcher as its entry point, because its slug carries a trailing word its function does not. It now publishes hybridProRataTime(incoming, price, lot, fraction, resting).

calculate is still exported from that subpath, so an import of it keeps working. What changed is run, which now aliases the topic's own function. If you called run("D12-F01-A04", inputs) there, call calculate instead — or better, the named function.

New#

• Financial Mathematics, Statistics, and Data Foundations — 120 topics, from means and percentiles through distributions, inference, regression, and risk and performance statistics. This is the base layer the rest of the library is meant to build on, so it lives at the short subpath foundations/…. It is deliberately absent from the README's algorithm index, which catalogues the market-facing topics; it ships, it is documented, and it is in the agent skill.
• Fundamental Analysis and Valuation — 52 topics. Discounted cash flow, dividend discount and Gordon growth, residual income, economic value added, relative valuation, and the distress and earnings-quality models: Altman Z, Piotroski F, Beneish M, Ohlson O, Dechow-Dichev, modified Jones. Plus integrated and sector-specific equity scoring.
• Model Validation and Backtesting — 10 topics. ROC and PR curves, Brier score, log loss, calibration and expected calibration error, gains and lift, cost-sensitive thresholds, score migration, slice validation, rare-event bounds.
• Credit Risk and Default — 7 topics. Logistic and probit PD models, through-the-cycle and point-in-time PD, Merton and Bharath-Shumway distance to default, Campbell-Hilscher-Szilagyi distress probability.
• Geometric chart patterns gains its continuation structures and pattern matching families — triangles, flags, pennants, wedges, DTW and matrix profile.
• 152 topics that could not ship now do. They carried an adapter rather than an implementation: a file that started a Python process, piped its input across as JSON and returned the answer. Correct, and unusable in a package that promises zero dependencies and has to run in a browser. The shared runtime behind them is now native TypeScript, and parity was measured rather than assumed — 298 full-series cases and three exact EPS cases compared against the Python reference observation by observation, 42,435 values, worst relative error 1.31e-15. This is what completes Technical Indicators, and it also brings the Hilbert and Ehlers cycle analytics, the pivot and market-structure families, and the last three EPS topics.

Verification#

The count nearly triples, and the description of it changes with it. These expected values are computed in the catalog by a Python implementation written alongside the TypeScript rather than derived from it, which makes the check cross-language parity, not an independent third-party figure. It catches transcription and generation errors — the failure mode that has actually occurred here — and it would not catch both implementations sharing a misreading of the source. The README and the agent skill now say so in those words.

Most of the jump comes from reading two fixture shapes the generator did not recognise. 102 topics already carried a separate input and expected-output pair on disk while nothing asserted their numbers, and the 152 ported topics carry a richer contract again: the warm-up index, every value in the final observation, and a causality property — that running over a prefix of the bars reproduces exactly the prefix of the full series, which is what catches an implementation that has quietly looked ahead. No expected value was authored to close either gap.

Contributing#

src/ is still generated and still overwritten on every sync, but there are now two routes for work that used to have none.

• optimised/ — a topic can ship a hand-written implementation from the repository instead of the catalog's. The catalog version is emitted beside it and a differential test asserts the two return identical values and throw identical errors, so an optimisation is reviewable by machine. npm run bench reports the ratio and CI fails if a recorded one regresses. The first is a 3.5x rewrite of calculateSma.
• A proposal route for new algorithms. Assigning a topic its id stays with the maintainer; everything after that is a normal pull request, and it can ship before the article exists.

See CONTRIBUTING.md.

Fixed#

• Ninety foundations topics would have published truncated subpaths — count-sum-minimu, look-ahead-leaka — because their folder names were capped at 28 characters and no slug was declared.
• Optional parameters were being dropped from the middle of published signatures, so a caller reading match(query, candidate, normalize) and passing three arguments was setting radius. Seven topics were affected.
• Arrow-form entry points recorded no parameters at all, publishing ten topics as zero-argument functions.
• engines.node is now >=22.12. The require condition resolves to an ES module and require(esm) is only unflagged from 22.12, so anyone on 22.0–22.11 was inside the declared range and getting a hard failure.
• The bundled lookup CLI reads the docs.json beside it before reaching for the network, instead of answering from a different version of the library.

Internal#

• A purity gate refuses any implementation that reaches a node: builtin, directly or through a shared module, and it now checks the whole import graph rather than one level. It is what caught the 152 adapter-backed topics before they could ship; they were ported rather than exempted, and the gate stays.
• gen-docs.mjs counts registry entries independently of the regex that parses them. Three times a topic had failed to match on one malformed field and vanished from the payload with no symptom.
• Relative imports that climb out of a topic directory are vendored into src/_shared/ generically, replacing a hardcoded case per engine.

0.12.0 — 2026-08-06#

The library gains no algorithms in this release and loses none: 324 topics before, 324 after. What changed is the surface in front of them. Every entry point now has one spelling, every subpath exports exactly its own function, and every topic carries a validated input contract.

If you are upgrading from 0.11.0, read the two tables below. 23 functions were renamed and 110 re-exports were removed; both fail loudly at import time rather than silently at run time, but both fail.

Breaking — entry points renamed to camelCase#

23 entry points were still snake_case, inherited from the reference implementations they were generated from. TypeScript consumers had to remember which of the 324 were the exceptions. They are now camelCase like the other 301.

No other entry point changed. The 301 that were already camelCase are untouched.

Breaking — subpaths no longer re-export their siblings#

28 subpaths used to re-export the other functions in their family: importing .../momentum/rsi also gave you williams_r, connors_rsi, stochastic_rsi and ultimate_oscillator. 110 such re-exports have been removed.

Why. A subpath is meant to be one function. Re-exporting siblings meant the same symbol had several valid import paths, so no import told you which module you actually depended on, and every family member was pulled in whether or not it was used.

The fix. Import each function from its own subpath — the third column of the table above. The symbol did not disappear; it stopped being in more than one place.

Every removed re-export is one of the 23 renamed functions, so a call site touched by this is usually touched by the rename too:

Added — CommonJS can resolve the package#

Every entry in the exports map now declares a require condition alongside import. Both point at the same ES module: there is no second build, the tarball is the size it was, and there is no dual-package hazard because there is only ever one copy of a module's state.

const { calculateEma } = require("fintech-algorithms/technical-indicators/trend-smoothing/ema");

Previously this failed with ERR_PACKAGE_PATH_NOT_EXPORTED — resolution, not loading, was the blocker: import is not a condition require() will match. require() of an ES module needs a runtime that supports it, which recent Node 22 does; the package's engines floor is unchanged at node >= 22. Resolves #1.

Added — the payload describes the surface that serves it#

docs.json (also published at /reference/payload.json) moves to schema 2.0.0. A consumer holding only the payload can now find the machine endpoints and map its own data onto the library without reading prose.

• package gained type, engines, license, the package's own languages (TypeScript, and only TypeScript), an entryPoints block stating that the root export is metadata-only and algorithms are subpath-only, and a machineSurface block naming /llms.txt, /version.json, /reference/payload.json and the per-domain and per-topic route patterns. These fields flow into /version.json unchanged.
• archetypes is new: the five input shapes as data — topic count, the canonical TypeScript input type, a runnable minimal payload, the validator to run at the boundary, and the one mistake each shape invites. This is the machine half of the archetypes guide.
• domains[].slug is new, so the URL for a domain no longer has to be rediscovered by finding one of its topics and splitting the path.

Changed — languages is now catalogLanguages (breaking for payload consumers)#

Every topic declared languages: ["typescript", "python"]. That is true of the catalog the package is generated from and false of the package, which ships TypeScript alone — and an agent reading it would confidently write pip install fintech-algorithms, which does not exist.

The per-topic field is renamed to catalogLanguages and is now read from the catalog's actual implementations/ directories rather than hard-coded. What this package ships is stated once, as package.languages.

Improved#

Every api block is checked against the implementation it documents at generation time — a declared parameter list that disagrees with the function's real signature fails the build. Contract coverage is now enforced in CI, so a topic can no longer arrive without one.

Internal#

• The generator refuses to emit a payload whose own structure is inconsistent — a domain spanning two path roots, or an archetype with no description.
• gen-docs.mjs takes --strict-contracts.
• CI smoke-tests the packed tarball from CommonJS as well as ESM.
• CHANGELOG.md ships inside the tarball.

0.11.0 and earlier#

Not recorded. docs.json inside each published version is the reliable record of what that version contained; the payload for any release is available at https://unpkg.com/fintech-algorithms@<version>/docs.json.