fix: correct ATR divergence, the EMA seed, and RSI on a flat series

[phmatray]

Three defects found while validating the samples in #104. All three are in released packages, and all three are the same species: a faithful-looking port of the TA-Lib C source that dropped exactly one line.

None of them could have been caught by the existing tests, which assert only RetCode == Success.

1. Atr never normalised its accumulator

src/TechnicalAnalysis.Functions/Atr/TAFunc.cs — main output loop divided on the way out but carried the undivided value into the next bar:

prevATR *= optInTimePeriod - 1;
prevATR += tempBuffer[today];
outReal[outIdx] = prevATR / optInTimePeriod;   // state keeps the undivided value

So the running average was multiplied by period - 1 every bar. On a series whose true range is a constant 4.0, where ATR(14) must stay exactly 4.0:

output	0	1	2	3	4
before	4.2857	4.4082	57.59	748.98	9737.02
after	4.0	4.0	4.0	4.0	4.0

Ratio is exactly 13.0 = period - 1. On 1500 bars, 1209 of 1486 outputs were +Infinity.

The warm-up loop directly above it, and the identical loop in Natr, both normalise. Only this one didn't.

2. TA_INT_EMA seeded itself low

Upstream while (i-- > 0) became i--; if (i <= 0) break; — which accumulates period - 1 values while still dividing by period. Testing the counter before decrementing rather than after is what loses the final term.

EMA(20) over a constant series of 100 returned 95.476190 instead of 100.

The seed is shared, so this reached Ema, Macd, MacdExt, MacdFix, Dema, Tema, T3, Apo, Ppo and Trix. Because the fast and slow legs are seeded with different periods, the error did not cancel: MACD read 0.53 on a series that never moved.

3. Rsi divided by zero on a flat window

A window with no price change has an average gain and an average loss of exactly zero, so 100 * gain / (gain + loss) was 0/0 — every element NaN, returned next to RetCode.Success and a non-zero element count. A halted instrument reaches this.

Upstream guards the same division and yields 0. All three division sites now route through one guarded helper, including the Metastock branch, which had the same unguarded divide.

Verification

The important part, given how these survived: every new test was confirmed to fail against the previous implementation.

with the fix:     Failed: 0,  Passed: 232
without the fix:  Failed: 6,  Passed: 226

The six failures are exactly the six regression tests. The seventh new test — RSI over a monotonically rising series — passes both ways by design: it is a control pinning the legitimate zero-loss case at 100, so the new zero guard cannot quietly swallow it.

All 225 pre-existing tests still pass, unchanged.

Follow-up

docs/guides/getting-started.md#10-known-library-defects and the sample guides in #104 document all three as known defects. Once this merges, those sections should be removed or rewritten — I left them alone here to avoid a conflict with #104. Merge #104 first, then this.

🤖 Generated with Claude Code