Appearance

Personalize your experience

Mode

Accent Color

Direction

Building Custom Indicators

Coding

Composing new indicators from existing ones, the recursive shapes that need care, warm-up handling, and how to check a custom indicator is actually correct.

Most custom indicators are compositions

Before writing anything, check whether the idea is already expressible. A surprising share of “custom indicators” are arithmetic over two built-ins, and arithmetic is available directly in a condition.

no custom code required

# oscillator spread — momentum divergence between horizons
RSI(14) - RSI(28)

# normalised volatility — comparable across instruments
ATR(14) / Close * 100

# Bollinger %B — where price sits inside the bands
(Close - BB_Lower(20,2)) / (BB_Upper(20,2) - BB_Lower(20,2) + 0.0001)

# volatility regime — is volatility expanding?
ATR(14) / ATR(14)[20]

# trend quality — slope measured against noise
LinRegSlope(Close, 20) / StdDev(Close, 20)

★

Pro tip:Compose first. A composition inherits correct warm-up handling, correct bar alignment and correct theming from its parts, and there is nothing to get wrong. Write a genuine custom indicator only when the calculation cannot be expressed as arithmetic over existing series.

The shape of an indicator

An indicator is a function from one or more series plus parameters to a new series of the same length. Two obligations come with that: the output must align bar-for-bar with the input, and the warm-up region must be marked rather than with a plausible-looking number.

the contract

1type Series = number[];23/**4 * Every indicator obeys three rules:5 *   1. output.length === input.length      (bar alignment)6 *   2. output[i] depends only on input[0..i]  (causality)7 *   3. output[i] is NaN while there is insufficient history (warm-up)8 */9type Indicator = (source: Series, ...params: number[]) => Series;1011const sma: Indicator = (source, period) =>12  source.map((_, i) => {13    if (i < period - 1) return NaN;                     // rule 314    let sum = 0;15    for (let k = i - period + 1; k <= i; k++) sum += source[k];  // rule 216    return sum / period;17  });                                                    // rule 1 by construction

Caution:Rule 2 is the one that quietly breaks. Any calculation that reads source[i + 1] — a centred average, a “smoothed” series computed in both directions, a peak detector that confirms with the next bar — is look-ahead bias. It will not error. It will produce the best backtest you have ever seen.

Recursive indicators

and anything built on it are recursive: each depends on the previous output rather than on a . That makes them cheap — constant work per bar — and introduces one subtlety worth understanding: the seed.

ema, and why the seed matters

1const ema = (source: Series, period: number): Series => {2  const k = 2 / (period + 1);          // smoothing factor3  const out: Series = [];4  let prev = NaN;56  source.forEach((value, i) => {7    if (i < period - 1) { out.push(NaN); return; }       // warm-up89    if (Number.isNaN(prev)) {10      // seed with the SMA of the first full window, not with source[0]:11      // seeding from a single bar makes the early output depend heavily on12      // one arbitrary value, and that bias takes ~3× period bars to decay13      let sum = 0;14      for (let k2 = i - period + 1; k2 <= i; k2++) sum += source[k2];15      prev = sum / period;16    } else {17      prev = value * k + prev * (1 - k);18    }19    out.push(prev);20  });2122  return out;23};

Seeding matters more than it looks. An seeded from a single bar carries that bar's influence for roughly three times the before it decays below noise. On a short backtest that region can be a meaningful share of the .
Recursive indicators never fully forget. Unlike a windowed average, an retains a decaying trace of all history. Two backtests over different start dates will produce slightly different EMA values on the same bar — which is a legitimate reason for two runs to disagree.
Cost is constant per bar. An does the same work whether the is 9 or 900, where an SMA does proportionally more. On long backtests with long periods that difference is measurable.

A worked custom indicator

Something not in the and not expressible as simple arithmetic: an efficiency ratio measuring how directly price travelled, which distinguishes a clean trend from a volatile that ended in the same place.

efficiency ratio

1/**2 * Efficiency ratio — net movement divided by total path travelled.3 *4 *   1.0  price moved in a straight line5 *   0.1  price wandered ten times as far as it progressed6 *7 * The same net change can be a clean trend or a noisy chop, and a8 * trend-following strategy wants to know which. Kaufman uses this9 * ratio to adapt smoothing; it also works as a standalone filter.10 */11function efficiencyRatio(close: Series, period: number): Series {12  return close.map((_, i) => {13    if (i < period) return NaN;1415    const net = Math.abs(close[i] - close[i - period]);1617    let path = 0;18    for (let k = i - period + 1; k <= i; k++) {19      path += Math.abs(close[k] - close[k - 1]);20    }2122    // a completely flat window has zero path; guard the division23    return path === 0 ? 0 : net / path;24  });25}

ℹ

Note:Read the guard on the last line. A where price did not move at all gives a zero denominator, which happens on halted instruments and illiquid far more often than it sounds. Returning 0 — “no ” — is the honest answer; returning NaN would also be defensible, but returning infinity is not.

using it as a filter

# only take trend entries when price is actually trending, not chopping
AND {
  EMA(21)  >  EMA(55)
  Close  crossesAbove  EMA(21)
  EfficiencyRatio(20)  >  0.3
}

Verifying correctness

A custom indicator that is subtly wrong is worse than none, because every built on it inherits the error. Four checks catch nearly everything.

Check	Type	Description
Length	structural	Output length equals input length. A shorter output silently misaligns every bar after the offset, so a condition compares this bar’s price against a different bar’s indicator.
Warm-up	structural	The first N−1 values are NaN, not zero. Zero is a number a comparison can succeed against, which produces in a region where the indicator has no meaning.
Known values	numerical	Compute a few values by on a short series and . Tedious once, and it catches off-by-one boundaries that no amount of inspection will.
Prefix stability	causality	Run it on the first 100 bars, then on the first 200, and the first 100 outputs are identical. If they differ, the indicator is reading the .

Practical guidance

Keep parameters few

Every parameter on a custom indicator is a parameter you will tune, and every tuned parameter is a of freedom fitted against your test . An indicator with four parameters offers far more ways to fit noise than it offers insight.

Normalise the output

An indicator returning price units is not comparable across instruments, so any threshold you set on it is -specific. Prefer output that is a ratio, a or a Z-score — then one condition works everywhere.

Prefer boring over clever

A composition of two built-ins that you fully understand beats a novel calculation you do not. The edge in a almost never comes from the indicator being unusual; it comes from the conditions around it being sound.

Caution:A custom indicator is also a maintenance liability. It has no reference implementation to check against, no community expectations about its behaviour, and nobody but you knows what it should do at the boundaries. Weigh that against what it actually adds before committing to one.

Appearance

Building Custom Indicators

Most custom indicators are compositions

The shape of an indicator

Recursive indicators

A worked custom indicator

Verifying correctness

Practical guidance

Keep parameters few

Normalise the output

Prefer boring over clever

See also

Related articles

Appearance

Most custom indicators are compositions

The shape of an indicator

Recursive indicators

A worked custom indicator

Verifying correctness

Practical guidance

Keep parameters few

Normalise the output

Prefer boring over clever

See also

Related articles

Related to Building Custom Indicators