Heikin Ashi Candles: The Exact Formula, and What the Smoothing Costs
Heikin ashi is a transform, not a data feed. It takes the same open, high, low and close every candlestick chart uses and rebuilds each bar out of averages, so the chart that comes out is smoother than the market that went in. Trends read as long single-colour runs. Chop mostly disappears.
That smoothness is bought with real information, and the price of it can be measured. This page gives the exact recurrence our chart engine runs, settles the repaint argument with a count rather than an opinion, and shows — on real Binance bars — how far the printed heikin ashi close sits from any price you could actually have traded.
What a heikin ashi chart shows
Every bar on the chart answers a different question than a raw candle does. A candlestick reports what price did inside one session. A heikin ashi bar reports where the session sat relative to a running average of the bars before it.
What you see is fewer colour changes. Across fourteen symbols and a thousand daily bars each, raw candles flipped colour 502 times per thousand bars — essentially a coin toss — while their heikin ashi transforms flipped 244 times. Mean run length doubled, from 1.99 candles to 4.09. Same tape, half the noise, and the chart below shows what that looks like when it works: one eleven-bar run covering a 23% move that the raw candles interrupted with red five times.
The exact heikin ashi formula our engine runs
heikin_ashi_bar runs. The close reads only the current raw bar; every trace of history arrives through the open, which is why colour persists and why the effect of the seed halves with each bar.Four lines, one of which does all the work.
How Vike computes it.heikin_ashi_barinvike-chartsets the HA close to(open + high + low + close) / 4of the raw bar, and the HA open to the midpoint of the previous HA bar's open and close — seeded on the very first bar with that bar's own(open + close) / 2. High and low then stretch to cover the body:max(high, ha_open, ha_close)andmin(low, ha_open, ha_close). One output bar per input bar, timestamps untouched.
Notice the close is a plain average of the current bar, so it carries no memory at all. Memory lives entirely in the open: each HA open is the previous HA bar's midpoint, which was itself built from the one before, and so on back to the seed. That recursion is why colour persists — an HA bar stays green as long as the four-price average keeps landing above a midpoint that only moves half the distance per bar.
I ported that rule from the crate source and checked it against the Python implementation the Rust names as its twin — fourteen symbols, a thousand bars apiece, all four prices compared on every one: 56,000 values, largest absolute difference 0.000e+00. Bit-identical, which is what makes the measurements below worth reading.
Does heikin ashi repaint? The seed, measured
Two separate claims hide inside the repaint accusation, and they deserve opposite verdicts.
Closed bars never change. The recurrence reads only the previous HA bar and the current raw bar, so arriving data cannot reach backwards: I recomputed the transform over the first 600 bars of each symbol and compared them against the same 600 inside the full series — zero of 8,400 bars differed, on any price. Whatever a heikin ashi bar looked like when it closed, it looks like forever.
Where your chart starts is another matter. Start the series 500 bars later and the seed differs, so the first bars disagree with the long chart: by a median 1.22% of price on the restart bar itself, and 5.76% in the worst case. The disagreement halves every bar — 0.34% two bars in, 0.04% by bar five, one part in ten million by bar twenty. Two heikin ashi charts of different history lengths converge fast, but they are genuinely different charts for their first dozen bars, and I've watched that difference get mistaken for a broken indicator more than once.
Reading heikin ashi candles
The standard playbook reads three things off each bar: colour for direction, body size for pressure, and the wick pattern for hesitation. Long green bodies with no lower wick mean the average keeps rising through its own midpoint — established uptrend. Small bodies with wicks both sides mean the bar closed near a midpoint that stopped moving, which is this chart's version of a doji.
Held against the raw tape, the readings mostly keep their promises about the past. The flip comes quickly too: when a heikin ashi run ends, the raw candles had typically flipped just one bar earlier (median lag 1 bar, at worst 4). What the smoothing changes is not when the turn shows up so much as how many false turns you no longer see.
Both charts above are the honest version of the pitch — a real trend, drawn twice, with the flip counts side by side.
The price on the chart is not a price
Here is the part I wish someone had shown me with numbers earlier.
An HA close is an average of four prices, so it is almost never a price the market printed at the close. Across 14,000 daily bars the gap between the heikin ashi close and the real close was 1.06% at the median, 4.26% at the ninety-fifth percentile, and 20.97% at the extreme. Every backtest that fills orders at HA prices inherits those gaps as free, imaginary edge.
That same illusion inflates finished trends. Over 1,140 single-colour runs of five bars or more, measuring each run from HA open to HA close showed a median 16.0% travel while the real close-to-close move was 15.2%. Not enormous — but it is a systematic flattering, and it compounds with the entry-price gap in the same direction. When I evaluate anything on this chart type, the rule is: signals may come from HA bars, but every fill and every stop is priced on the raw tape.
Where heikin ashi fails
Smoothing is not protection, and the failure has a shape.
Of 3,423 heikin ashi colour runs in the sample, 39.8% were over within two bars. The whipsaw below is one of 731 cases where a fresh flip was followed by a move of eight per cent or more in the opposite direction within five bars — the LINK flip lasted two bars and cost 9.1% to anyone who treated it as an entry. A calmer chart does not mean a calmer market; it means the chart tells you about the market slightly later and with fewer words.
A second failure mode is subtler: candlestick patterns stop meaning anything. A hammer is defined on real session prices — its long lower shadow is a claim that sellers pushed and were rejected at specific levels. Run the transform and the shadow you see belongs to synthetic prices; our pattern detectors read raw OHLC only, and a "hammer" spotted on heikin ashi bars is a shape the market never made.
Heikin ashi vs candlesticks: which to use
Use them for different jobs. Raw candles are the record; heikin ashi is a summary, and summaries are for context rather than execution.
One combination survives contact with practice: judge the trend state on the transform, then act on raw prices — pattern, level, stop and fill all on the true tape. Pairing the smoothed chart with an independent trend tool like the ichimoku cloud also works, precisely because the two disagree at different moments. What I skip entirely is reversing a position because one HA bar changed colour: 658 of those runs lasted exactly one bar.
Frequently asked questions
Does heikin ashi repaint? Closed bars, never — zero changed values in 8,400 recomputed bars. But charts with different history lengths disagree near their start (median 1.22% at the seed, gone within about twenty bars), because the seed is the first bar's own open/close midpoint.
Can I trade directly off heikin ashi signals? You can enter off them, but you cannot price off them. The HA close sits a median 1.06% from the real close on daily crypto bars, and up to 21% in the tail. Fills and stops belong on raw prices.
Why do my heikin ashi values differ from another platform's?
Usually the seed, occasionally the history length. Our engine seeds the first HA open with (open + close) / 2 of the first bar; platforms that seed with the raw open alone produce bars that disagree until the recursion washes the difference out.
Do candlestick patterns work on heikin ashi charts? No — the shapes are made of averaged prices, so a pattern's premise (rejection at real levels) is gone. Detect patterns on raw candles, use the transform for trend context.
This is educational material, not financial advice. Every figure above was measured on past data, and a smoother chart never removes the risk in the market underneath it — size positions so a wrong read costs little.