Bollinger Bands

Where is the current close relative to its rolling level and dispersion? That is the practical question behind Bollinger Bands. This tutorial builds the answer from finalized data, shows the exact arithmetic, and gives you a laboratory for breaking the indicator safely.

The mental model in one minute

This package uses the traditional close-price SMA and population standard deviation over the same 20-row window, with symmetric bands at two standard deviations.

The output is a measurement. It is not an order, forecast, confidence interval, or promise that a familiar-looking market pattern will repeat.

Contract before chart

Variant: This package freezes close, SMA(20), population standard deviation, symmetric multiplier 2, and no display offset.
Defaults: period=20, multiplier=2.
First publication: index 19 after the first 20 closes.
Data: finalized, causal observations on one price basis.
Missing, malformed, or mixed-basis input: reject rather than repair silently.

Formula and symbols

$Middle_t=SMA_n(C)_t,\quad Upper/Lower_t=Middle_t\pm k\sigma_{pop,t}$

Symbol	Meaning	Unit
$C_i$	close inside the active window	price units
$M_t$	n-period simple moving average	price units
$σ_t$	population standard deviation of the same n closes	price units
$k$	positive band multiplier	dimensionless
$\%b_t$	location between lower and upper bands	ratio

Calculate it by hand

Use the complete four-close window [10, 12, 14, 16] and k=2.

Mean: (10 + 12 + 14 + 16) / 4 = 13.
Squared deviations: 9, 1, 1, 9; total 20.
Population variance: 20 / 4 = 5; deviation √5 ≈ 2.236068.
Upper: 13 + 2 × 2.236068 ≈ 17.472136.
Lower: 13 - 2 × 2.236068 ≈ 8.527864.
For the current close 16, %b ≈ (16 - 8.527864) / 8.944272 ≈ 0.835410.

Using sample deviation would divide by 3 and create wider bands. That is a different contract, not a rounding difference.

That calculation is deliberately small enough to audit without code. The 240-row fixtures then show how the same rule behaves through warm-up, regime changes, ties, gaps, shocks, and expiry.

The complete algorithm

Take the trailing n closes including the current close.
Compute their arithmetic mean.
Compute population variance with denominator n, then its square root.
Add and subtract k times that deviation from the same mean.
Calculate %b only when upper minus lower is nonzero.

Rendering system map…

The flow exposes the point where a result becomes unavailable, where recursive state begins, or where a rolling owner expires.

See the mechanism

Mechanism map for Bollinger Bands

The map connects formula terms to the exact data they consume.

Follow a complete causal trace

Canonical 240-row trace for Bollinger Bands

Only rows at or before the active observation may influence the displayed state. Warm-up is shown as unavailable rather than silently converted to zero.

Memory is part of the definition

Finite n-row window. An outlier changes mean and dispersion immediately, then can cause an abrupt width change when it leaves the back of the window.

Band width equals 2kσ. %b may be below 0 or above 1. A zero-variance window collapses all bands and makes %b undefined.

Memory and family comparison for Bollinger Bands

This is why two indicators can react differently to the same shock even when their plotted bands appear similar.

Six ways to try to break it

Scenario	Teaching job	Expected lesson
`contraction-expansion`	Contraction then expansion	Population dispersion narrows, then widens.
`band-walk`	Trending band walk	Repeated upper-band proximity does not imply reversal.
`population-audit`	Population versus sample audit	The denominator is n, not n minus one.
`zero-variance`	Zero-variance window	Bands collapse and percent-b is unavailable.
`outlier-expiry`	Single-outlier entry and expiry	A one-row outlier widens the finite window, then leaves it abruptly twenty rows later.
`responsive`	Responsive 10 × 1.5 bands	Shorter, narrower settings react faster.

Boundary atlas for Bollinger Bands

Predict the active component before opening the lab. Then jump to the relevant checkpoint and inspect the formula trace rather than guessing from line shape.

Interactive laboratory

Open the Bollinger Bands guided lab. It contains six 240-row scenarios. Change the period or multiplier where the formula exposes one; the lab recomputes immediately, stops playback, resets to an informative state, and preserves the same calculation path used by Step. Use the checkpoint selector to compare first publication, mechanism, boundary, memory, platform reconciliation, and mature state.

Production implementation

The Python and TypeScript references are topic-scoped and intentionally readable. They validate before calculating, return aligned arrays, and use None/null for unavailable states. Shared fixtures establish parity; independent family checks establish properties without trusting either implementation as the oracle.

When production output disagrees with a chart, log:

symbol and source ownership;
timeframe, session, timestamp, and bar finality;
adjustment basis;
parameter values and first-ready index;
the earliest divergent component;
window inclusion, seed, denominator, multiplier, equality, and null policy;
the earliest revised observation.

Compare, do not conflate

Nearby topic	What changes
Keltner Channels	use recursive EMA level and ATR distance instead of rolling close dispersion.
Donchian Channels	use extrema rather than a mean and dispersion.
BandWidth	normalizes this absolute band width by the middle band.

Misuse boundary

Do not: Treating two standard deviations as a 95% confidence interval for prices.
Do not: Treating an upper-band tag as an automatic sell or a lower-band tag as an automatic buy.
Do not: Mixing sample and population deviation or using different windows for mean and deviation.

The synthetic examples establish implementation behavior only. No historical association, transaction cost, out-of-sample result, or investment suitability claim is made.

Historical example decision

No named historical episode is used. The mechanism is learned more cleanly from controlled synthetic paths, and a named case would require point-in-time, licensed, adjustment-consistent OHLC evidence before publication.

Summary

You can now state the precise question Bollinger Bands answers, calculate it, explain its memory, diagnose a platform mismatch, and separate a measurement from a trading claim. Use the family comparison rather than chart resemblance to choose the next concept.

Primary evidence

See the topic reference ledger for source roles, dates, and limitations.

Workflows

Rendered from the canonical Mermaid sources linked by this article.

Bollinger Bands calculation flow

Purpose: show the exact selected state, readiness branch, and output.

Rendering system map…

Takeaway: the formula, memory, and unavailable state are part of the contract, not chart formatting.

ReferencesPrimary sources and evidence notes

Expand the source trail, evidence role, and limitations behind the engineering choices.

Bollinger Bands Rules

Organization or authors: John Bollinger
Source type: Creator-maintained guidance
Publication, version, or access date: site current at access 2026-07-26
URL or identifier: Bollinger Bands Rules
Accessed: 2026-07-26
Jurisdiction: not jurisdiction-specific
Evidence role: sourced definition, maintained implementation, or bounded historical context
Supports: Supports default-versus-universal distinctions, band-tag warnings, traditional SMA consistency, non-normality warning, %b, and BandWidth relationships.
Limitations: Interpretive guidance is not evidence of prediction or profitability.

Bollinger Bands (BB)

Organization or authors: TradingView
Source type: Maintained platform documentation
Publication, version, or access date: accessed 2026-07-26
URL or identifier: Bollinger Bands (BB)
Accessed: 2026-07-26
Jurisdiction: not jurisdiction-specific
Evidence role: sourced definition, maintained implementation, or bounded historical context
Supports: Documents the common 20-period SMA and two-deviation presentation and configurable inputs.
Limitations: Platform behavior is not a universal mathematical standard.

BBANDS — Bollinger Bands

Organization or authors: TA-Lib
Source type: Maintained official documentation
Publication, version, or access date: maintained; accessed 2026-07-26
URL or identifier: BBANDS — Bollinger Bands
Accessed: 2026-07-26
Jurisdiction: not jurisdiction-specific
Evidence role: sourced definition, maintained implementation, or bounded historical context
Supports: Supports the moving-average and standard-deviation band structure.
Limitations: TA-Lib supports middle-band variants beyond this package’s SMA.

ta_BBANDS.c

Organization or authors: TA-Lib
Source type: Maintained reference implementation
Publication, version, or access date: main branch; accessed 2026-07-26
URL or identifier: ta_BBANDS.c
Accessed: 2026-07-26
Jurisdiction: not jurisdiction-specific
Evidence role: sourced definition, maintained implementation, or bounded historical context
Supports: Supports executable band construction and deviation handling.
Limitations: Library configuration is broader than the frozen lesson variant.

e-Handbook: Measures of Scale

Organization or authors: NIST/SEMATECH
Source type: Government statistical reference
Publication, version, or access date: accessed 2026-07-26
URL or identifier: e-Handbook: Measures of Scale
Accessed: 2026-07-26
Jurisdiction: not jurisdiction-specific
Evidence role: sourced definition, maintained implementation, or bounded historical context
Supports: Distinguishes standard deviation as a scale measure and supplies statistical context.
Limitations: It does not define a financial indicator or validate trading interpretations.

Public evidence boundary

The fixture is author-generated synthetic teaching material. Every displayed number is an author-derived calculation from those inputs. No provider observation, historical return, causal claim, forecast, or profitability result is published. Platform documentation supports only the named platform convention; local seeds, nulls, equality, and scope remain implementation choices.

Implementation

Full dependency-light reference implementations in both supported languages.

volatility_channels.ts

function finite(v:number[],n:string){if(!Array.isArray(v)||!v.length||v.some(x=>!Number.isFinite(x)))throw new Error(`${n} must contain finite numbers`)}
function hlc(h:number[],l:number[],c:number[]){finite(h,"high");finite(l,"low");finite(c,"close");if(h.length!==l.length||h.length!==c.length)throw new Error("high, low, and close lengths must match");if(h.some((x,i)=>x<l[i]||c[i]<l[i]||c[i]>x))throw new Error("each bar requires low <= close <= high")}


function period(v:number,n="period"){if(!Number.isInteger(v)||v<1)throw new Error(`${n} must be a positive integer`)}

function rolling(close:number[],p:number){period(p);finite(close,"close");const middle:(number|null)[]=Array(close.length).fill(null),stddev:(number|null)[]=Array(close.length).fill(null);for(let i=p-1;i<close.length;i++){const w=close.slice(i-p+1,i+1),m=w.reduce((a,b)=>a+b,0)/p;middle[i]=m;stddev[i]=Math.sqrt(w.reduce((a,b)=>a+(b-m)**2,0)/p)}return{middle,stddev}}
export function bollinger_bands(close:number[],p=20,multiplier=2){if(!Number.isFinite(multiplier)||multiplier<=0)throw new Error("multiplier must be finite and positive");const r=rolling(close,p),upper:(number|null)[]=Array(close.length).fill(null),lower:(number|null)[]=Array(close.length).fill(null),percent_b:(number|null)[]=Array(close.length).fill(null);for(let i=0;i<close.length;i++){if(r.middle[i]===null)continue;upper[i]=(r.middle[i] as number)+multiplier*(r.stddev[i] as number);lower[i]=(r.middle[i] as number)-multiplier*(r.stddev[i] as number);const width=(upper[i] as number)-(lower[i] as number);percent_b[i]=width===0?null:(close[i]-(lower[i] as number))/width}return{middle:r.middle,stddev:r.stddev,upper,lower,percent_b}}

Playground

The embedded lab now expands to its full document height, keeping the article as the only scroll surface.

Full-height labbollinger bands labOpen full screen