How formulas are verified

Every indicator in zeon-ta is implemented from a formula that has been checked against an independent, authoritative source — never from a single source trusted at face value, and never from memory. This page explains why, and what the process actually looks like.

Why this is a hard rule

Early in this project, two indicators were implemented from a single source’s stated formula and both turned out to be wrong once cross-checked against a second source:

• trend_channel’s bands were measuring how far price scattered around the window’s mean, when the standard definition (and the source’s own prose, read carefully) actually calls for scatter around the fitted regression line — a different, smaller number that inflates the channel incorrectly.
• squeeze’s momentum midline used a flat three-way average of its three inputs, when the canonical TTM Squeeze formula uses a specific nested average that weights two of them together first.

Neither mistake was a typo — both came from trusting one source’s wording without checking it against how the indicator is actually computed elsewhere. That is the whole reason the verification rule below exists, and why it applies to every indicator added since, not just the ones that broke.

The process

1. Find the formula in a primary, authoritative source. StockCharts ChartSchool is preferred — it is maintained, cites the original developer where one exists, and is precise about defaults and edge cases. When ChartSchool has no page for an indicator, the fallback order is: Fidelity’s Technical Indicator Guide, Wikipedia, or the originating platform’s own official documentation (MetaTrader5, TradingView) for indicators native to a specific platform.
2. Cross-check against a second source whenever the first is ambiguous about a default parameter, a rounding rule, how the very first bar is handled, or a division-by-zero edge case. Two sources agreeing on the core formula but differing on a default is common and worth noting in the docstring; two sources actively disagreeing on the formula itself is a sign to keep looking rather than pick one arbitrarily.
3. Record the source. The confirmed URL goes on the function’s @indicator(reference=...) and in its docstring’s References section, so anyone reading the code — not just this page — can check the implementation against the same source.
4. Compute every example, never guess it. A docstring’s Examples block is executed by the test suite, but the value written into the docstring is authored by a person first. Several examples in this library’s history were initially guessed and turned out to be wrong once actually computed — hma’s doctest was guessed as 30.0 on a straight ramp and was actually 29.3333, williams_r’s was guessed as 0.0 and was actually -0.0 (a genuine float sign edge case), stoch_rsi’s was guessed as 100.0 and was actually 50.0 (a flat-market convention triggered by RSI itself pinning at its own ceiling). None of these were caught by reasoning about the formula harder; all were caught by running the code before trusting the number.
5. Write tests that would have caught the two mistakes above. At least one golden value traced by hand against the confirmed formula, not against the implementation’s own output — a test that reproduces a bug instead of catching it is worse than no test.

What this does not cover

A handful of indicators (marked with lesson instead of reference in the registry) predate this project’s move to citing external sources per function; they follow the same standard, widely published definitions, just without an individually cited URL in their docstring. Every indicator added since — see CHANGELOG.md’s 0.2.0 entry onward — carries an explicit reference.

See CONTRIBUTING.md for how this applies when adding a new indicator.