ta.go 2.1.0 (tag v0.21.0)

Released 2026-10-10. Full details in CHANGELOG.md.

Install or pin this line with go get github.com/Bitvested/ta.go@v0.21.0; ta.Version reports 2.1.0.

⚠️ Read this before upgrading

Four functions return different numbers, and three of them a different number of rows. Each was verified against ta.js, the reference these ports are built from, and each fix has a test that fails against the old code — but nothing raises, so an existing pipeline will consume the new values silently.

function	what changed
HalfTrend	row count: n - atrlen → n. It also produced zero trend changes on real data before this release.
Alligator, Gator	row count: n - jl - js + 2 → n, oldest-first and aligned to the input bars.
Kama	row count n - l1 → n - l1 + 1, and every value — the smoothing constant had lost its slow term.
Hull	values: the half-length and the final smoothing window were truncated instead of rounded, so the output differed from ta.js at every odd length and wherever sqrt rounds up (8, 14, …).
ZigZag	values: three bookkeeping defects — legs were reset from ±Inf, the running maximum tracked the low, and a low pivot stored the high.

Er, Ar and Kelly are not in that list: their values are unchanged. What changed is that a one-sided series no longer depends on math.Pow(1, +Inf) happening to be 1.

Added

• Occ — the Open Close Cross, following JustUncleL's Pine (source), requested in Bitvested/ta.py#4. A moving average of open and of close; the direction is whichever is on top and a flip is the cross. Input rows are [open, close]; output rows are [close basis, open basis, direction], aligned to the input bars, NaN until both bases exist. delay is the Pine's delayOffset. basisType accepts SMA, EMA, DEMA, TEMA, WMA, SMMA, HullMA and TMA; the other four the Pine offers are omitted rather than approximated (LSMA needs an offset this package's Lsma does not take, VWMA needs volume in the row, ALMA and SSMA are not implemented). An unrecognised basisType returns nil.

Fixed

Fixed

• Alligator and Gator now align to the input bars. Both returned
  their rows newest-first and truncated to n - jl - js + 2 rows, so
  the output could not be lined up with the input series at all, and
  each line was displaced by shift - 1 bars instead of shift.
  Gator's upper histogram also plotted jaw - teeth instead of
  |jaw - teeth|, giving the top histogram the wrong sign. Both now
  emit one row per input bar, with each line displaced right by its own
  shift, matching TradingView's plot(..., offset=shift), and NaN until
  that line's first bar (jaw 20, teeth 12, lips 7 with the defaults).
  Gator is now derived from Alligator's displaced lines, and
  js=ts=ls=0 reproduces TradingView's Gator Oscillator, which does not
  displace its lines. Mirrors the same fix in ta.py and ta.js.

• Kama now computes Kaufman's adaptive average correctly. Four bugs,
  all inherited from the ta.js original: the smoothing constant dropped
  its slow term (- 2/(l3+1) + 2/(l3+1) cancelled, leaving ER * fast);
  the volatility window summed l1-1 one-bar changes instead of l1,
  which let the efficiency ratio exceed 1 and divided by zero on flat
  input; the seed was the SMA of the last window rather than the first,
  so the opening value was computed from data at the end of the series;
  and the loop started one bar late, emitting n - l1 values instead of
  n - l1 + 1.

  The recursion now matches tulind, the reference the benchmarks compare
  against: bit-identical on short, flat, trending and gapping series, max
  Δ 1.4e-14 on a 100k random walk. Carrying the volatility sum across
  bars instead of re-summing it per bar also makes it ~2.5× faster.
  Fixes the same defect as Bitvested/ta.js#22,
  mirrored in ta.py.

• HalfTrend no longer sits on one line forever. Reported against ta.py
  as #7. The port read the
  SMA at the start of its window — roughly atrlen bars old — instead
  of at the current bar, so both flip conditions compared stale values and
  never fired: on 4901 bars of 5-minute data it produced zero trend
  changes. The extremes behind maxlow/minhigh also came from the ATR
  window rather than amplitude bars, the down-flip compared close with
  < where the Pine uses >, and the window data[i-atrlen:i] excluded
  the current bar, leaving every value a bar stale.

  The rewrite follows everget's Pine
  line by line. Its ATR uses Pine's own seeding (Wilder's RMA seeded with
  the mean of the first atrlen true ranges, which is not what Atr
  does). Verified against a transcription of the Pine: exact agreement,
  160 flips where there were none, across the three ports.

• Er no longer leaves a one-sided series to an exponentiation edge case.
  It took the geometric mean of both sides regardless of how many
  observations each had, so a series with no losses relied on
  math.Pow(1, +Inf) happening to return 1 — which is not what the other
  ports did with the same input: ta.js returned NaN (1 ** Infinity is
  NaN in JS) and ta.py raised ZeroDivisionError. An empty side carries
  weight zero, so its mean is never taken and it contributes nothing; all
  three now state that explicitly and return identical numbers. Er and
  the Ar and Kelly built on it return 0.0274, and a Kelly fraction of
  1, for an all-winning series, and -1.0276 for an all-losing one. Mixed
  series are unchanged to the last bit. Mirrors the same fix in ta.js and
  ta.py.

• ZigZag's pivot bookkeeping matched neither ta.js nor its own documentation.
  Three defects: the leg reset seeded the running extreme with ±Inf instead of
  the pivot just recorded, so the next retracement was measured from nothing;
  the running maximum tracked the low rather than the high; and a low pivot
  stored the high. On a 300-bar series it disagreed with ta.js in 174 of 298
  values; it now matches exactly (max Δ 1.4e-14). The existing fixture used a 25%
  retracement, where the defects cancel out, which is why they survived — a 5%
  case now pins them.

• Hull truncated where the reference rounds. Both the half-length and the final
  smoothing window came from int(), so the output disagreed with ta.js at every
  odd length (half-length) and wherever sqrt rounds up — length 8 has sqrt 2.83,
  rounded to 3 by the reference and floored to 2 here, which also shifted the
  output onto a different bar grid. Both now use math.Round, and a sweep of
  lengths 2-20 matches ta.js exactly. Found while checking Occ's HullMA basis,
  which inherits it.

Changed — BREAKING for HalfTrend consumers

• HalfTrend returns one row per input bar instead of n - atrlen, so
  row i covers bar i. The ATR channel is NaN until the ATR window is
  full (bar atrlen - 1); the trend line itself is defined from bar 0.
  Callers that index-shifted the old output to align it with their data no
  longer need to.

v 0.20.0

Changelog

All notable changes to ta.go.

v2.0.0 — git tag v0.20.0

Major refactor. ~52 new functions, formula and convention fixes on
existing ones, no-mutation guarantees, a goroutine Monte Carlo
subpackage, and performance rewrites that make ta.go competitive with
or faster than markcheno/go-talib and cinar/indicator at 100 k
bars.

Version = "2.0.0" is exported from the root ta package; the Git
tag itself is v0.20.0 so the existing github.com/Bitvested/ta.go
import path keeps working without Go's /v2 module-path rule.

Layout

• The single ta.go file is gone. Implementations now live under
  category subpackages:
  • src/moving_averages (movingaverages)
  • src/indicators
  • src/oscillators
  • src/bands
  • src/statistics
  • src/chart_types (charttypes)
  • src/experimental
  • src/random (new package)
  • src/misc
• New top-level multi/ subpackage for goroutine Monte Carlo.
• New top-level bench/ Go module (separate go.mod) for
  cross-library comparative benchmarks — does not pollute the
  consumer go.sum.
• New internal/testutil with tu.Slice, tu.Slice2D,
  tu.Scalar — float-tolerant comparators (rel/abs tolerance
  1e-12, NaN-aware).

Formula fixes

• indicators.Psar — downtrend extreme now tracks the bar's
  low (data[i][1]); previous code mistakenly tracked the high.
• statistics.Kmeans — value-copies old, resets means per
  loop, recomputes by cluster index, and resets changed.
• indicators.Ichimoku — Senkou Span B's lows window now uses
  length3; the per-window max/min loops were replaced with
  rolling-extremum scans.
• movingaverages.Smma — Wilder recurrence with SMA seed of
  the first N values; output length is N − l + 1.
• indicators.Alligator / oscillators.Gator — fixtures
  regenerated against the canonical Williams formulas (cascades
  from the Smma fix).
• oscillators.MomOsc — Chande CMO formula on first
  differences (positive vs absolute negative deltas).
• indicators.Fractals — returns [][]bool; the broken
  price-mode shape was replaced with a separate
  indicators.FractalsPrice that pads two leading and two
  trailing [-1, -1] rows.
• indicators.ZigZag — fixed the state-transition tracker
  (the scalar branch was resetting the wrong extreme). Both
  scalar and OHLC branches share consistent semantics.

Convention changes (breaking)

• indicators.Rsi — Wilder smoothing (TA-Lib convention) is
  the new default. The legacy simple-average behaviour is exposed
  as Wrsi for back-compat callers.
• indicators.Roc — returns percent (×100). Cop and Kst
  fixtures regenerate as a linear cascade.
• indicators.Mfi — input is [][]float64 of [H, L, C, V]
  bars (was pre-computed positive/negative flow).
• statistics.Median — even windows return the average of the
  two middle values (was lower-median). Mad cascades.
• indicators.Atr — output length is N − l + 1 (was N); TR
  uses absolute values for |H − Cprev| and |L − Cprev|. Keltner,
  Supertrend, Halftrend lose their compensating splice.
• oscillators.Fisher — output length is N − l + 1 (drop
  trailing slice).

No-mutation guarantees

The following functions no longer mutate caller inputs:
statistics.Std, indicators.Cross (now also returns
[]CrossSignal{Index, Cross}), indicators.Divergence,
statistics.Percentile, experimental.Support / Resistance.

NaN handling standardised

Rsi, Roc, Dif, Mom (percent mode), MomOsc, Mfi,
Fisher, Pr, Stoch now return math.NaN() on flat-window /
divide-by-zero edges instead of silent 0/1.0 coercions.
statistics.Kelly no longer coerces NaN → 1.0;
indicators.ChaikinOsc no longer coerces → 0.

New functions (~52)

Moving averages (6)

Dema, Tema, T3, Trima, Vidya, Zlema.

Indicators (26)

Adl, Adx, Adxr, Apo, Cci, Cmf, Cross, Dpo, Dx,
Emv, Kdj, Mass, Mdi, Mdm, Natr, Nvi, Pdi, Pdm,
Ppo, Pvi, StochRsi, Trix, Ulcer, Vortex, plus the
Macd split: MacdSignal, MacdBars.

Oscillators (2)

Kvo, Ult.

Statistics (9)

Covariance, ExpectedTrails, LrSlope, LrIntercept,
LrAngle, Tsf, StdSeries, ReturnPositive, ReturnNegative.
(Ssd already existed.)

Random (5, new package)

Prng (deterministic mulberry32 closure), Pick, Range,
Float, Order. Sampling helpers accept an optional rng with
math/rand fallback when nil.

Experimental (1)

DivergenceState.

While porting DivergenceState, two latent bugs in
experimental.Support were fixed: the in-package minf/maxf
helpers sorted the caller's slice in place (now copy first), and
the anchor-advance branch set findex = true instead of false
(so the outer do-while exited after one iteration). Resistance
was already correct on the second point.

Performance rewrites

Captured per-rewrite in bench/results-phase4.txt. Headline
speedups at 100 k bars:

Function	Before	After	Speedup
Sma	30 ms	154 µs	≈195×
Variance	110 ms	182 µs	≈600×
Bands	22 ms	524 µs	≈42×
Ema	960 µs	320 µs	≈3×
Wma	800 µs	230 µs	≈3.5× (10 k)
Rsi	935 µs	520 µs	≈1.8×

Plus drop-per-window-copies/sort in AroonUp/AroonDown/
Elderray/Don/Fisher (≈13–50× at 10 k).

Multi-thread Monte Carlo

New multi/ subpackage:

• multi.Sim(data, length, sims, perc) []float64 — fans
  min(sims, runtime.NumCPU()) worker goroutines reading from a
  job channel; returns data followed by the per-bar perc-rank
  quantile of the ensemble.
• multi.SimSeeded(data, length, sims, perc, seed) []float64 —
  deterministic variant. All uniform draws are pre-generated in a
  single goroutine before fan-out, so output depends only on
  (data, length, sims, perc, seed).

Cross-library benchmarks

New bench/ separate Go module compares ta.go against
github.com/markcheno/go-talib and github.com/cinar/indicator
for SMA, EMA, RSI, MACD, Bollinger Bands, and ATR at 1 k / 10 k /
100 k bars. ta.go ties or wins at 100 k on every benched
indicator. See bench/BENCHMARKS.md.

Tests

Test count grew from 99/100 (one pre-existing float-drift
failure) to 165/165 passing. Float-tolerant comparator at
internal/testutil/close.go underpins fp-sensitive fixtures.

Full Changelog: 0.0.14...v0.20.0

v 0.0.13

Full Changelog: 0.012...0.0.13

v 0.0.12

Full Changelog: 0.0.11...0.012

v 0.0.11

Full Changelog: 0.0.10...0.0.11

v 0.0.10

Full Changelog: 0.0.9...0.0.10

v 0.0.9

Full Changelog: 0.0.8...0.0.9

v 0.0.8

Full Changelog: 0.0.7...0.0.8

v 0.0.7

Full Changelog: 0.0.6...0.0.7

v 0.0.6

Full Changelog: 0.0.5...0.0.6