Skip to content

Streaming / Stateful

tulip_rs indicators implement a from_state path: once the initial lookback has been computed, each subsequent bar is updated using only the saved state — no need to reprocess the full history.

All timings are nanoseconds (ns). Lower is better.

  • Batch (ns) — full indicator computation over the complete 6,705-bar dataset
  • Streaming 1-bar (ns) — time to update one new bar using saved state
  • Batch / Streaming — how many times faster the streaming path is vs a full batch recompute
Indicator Batch Rust (ns) Streaming 1-bar (ns) Batch / Streaming
ad 4,779 25 191.2×
adosc 6,556 27 242.8×
adx 9,999 80 125.0×
adxr 11,039 75 147.2×
ao 5,393 14 385.2×
apo 4,595 23 199.8×
aroon 17,988 38 473.4×
aroonosc 19,873 33 602.2×
atr 4,559 23 198.2×
avgprice 1,362 22 61.9×
bbands 7,114 34 209.2×
bop 2,312 24 96.3×
cci 55,726 40 1393.2×
chaikinmf 6,833 19 359.6×
chandelierexit 17,252 38 454.0×
cmo 5,959 29 205.5×
cvi 5,926 16 370.4×
dema 5,993 22 272.4×
di 9,713 109 89.1×
dm 6,611 66 100.2×
donchianchannel 15,182 36 421.7×
dpo 2,528 24 105.3×
dx 8,441 66 127.9×
ef 4,574 25 183.0×
elderray 6,197 25 247.9×
ema 4,547 14 324.8×
emv 2,295 24 95.6×
fisher 48,523 36 1347.9×
fosc 7,513 32 234.8×
hma 8,114 31 261.7×
kama 6,892 24 287.2×
keltnerchannel 6,182 33 187.3×
kvo 9,068 30 302.3×
linreg 6,228 29 214.8×
macd 7,045 21 335.5×
marketfi 2,309 24 96.2×
mass 5,568 18 309.3×
max 5,360 27 198.5×
md 12,120 28 432.9×
medprice 1,009 23 43.9×
mfi 7,513 27 278.3×
min 7,808 28 278.9×
mom 838 26 32.2×
msw 625,817 114 5489.6×
natr 4,712 26 181.2×
nvi 2,395 23 104.1×
obv 3,325 23 144.6×
ppo 4,823 25 192.9×
psar 10,055 25 402.2×
pvi 2,400 23 104.3×
qstick 2,731 27 101.1×
roc 2,294 26 88.2×
rocr 2,289 26 88.0×
rsi 4,719 69 68.4×
sma 2,451 24 102.1×
smaenvelope 7,266 31 234.4×
stddev 3,601 30 120.0×
stoch 20,391 40 509.8×
stochrsi 19,106 44 434.2×
tema 6,734 26 259.0×
tr 1,545 23 67.2×
trima 5,408 25 216.3×
trix 6,457 27 239.1×
trvi 6,010 19 316.3×
tsf 6,270 29 216.2×
typprice 1,128 14 80.6×
ultosc 16,033 28 572.6×
vhf 14,842 29 511.8×
vidya 11,607 35 331.6×
volatility 8,689 29 299.6×
vortex 7,504 28 268.0×
vosc 3,791 27 140.4×
vwma 3,310 30 110.3×
wad 3,751 23 163.1×
wcprice 1,058 22 48.1×
wilders 4,536 14 324.0×
willr 17,218 34 506.4×
wma 6,155 16 384.7×
zlema 5,716 24 238.2×

RustTa processes data one bar at a time (streaming), so its batch time is effectively its per-update cost × 6,705 bars. The comparison below shows tulip_rs streaming (single-bar) vs RustTa's full-batch equivalent for 20 indicators.

Indicator tulip_rs Streaming 1-bar (ns) RustTa Batch (ns) RustTa / tulip_rs
atr 23 9,342 406×
bbands 34 8,093 238×
cci 40 65,098 1627×
chandelierexit 38 42,341 1114×
ef 25 26,694 1068×
ema 14 7,919 566×
keltnerchannel 33 12,597 382×
macd 21 7,891 376×
max 27 17,229 638×
md 28 26,952 963×
mfi 27 23,766 880×
min 28 22,910 818×
obv 23 5,216 227×
ppo 25 7,900 316×
roc 26 2,860 110×
rsi 69 7,909 115×
sma 24 4,533 189×
stddev 30 8,346 278×
stoch 40 46,721 1168×
tr 23 7,502 326×
Key findings
  • All 79 indicators implement the stateful streaming path
  • Median streaming time: ~27 ns per bar update
  • Fastest — ao, ema, typprice, wilders at 14 ns/bar
  • Slowest — msw at 114 ns/bar
  • The streaming path is 100–5,489× faster than full batch recompute depending on the indicator
  • tulip_rs streaming is 100–1,627× faster than RustTa's effective per-update cost across 20 indicators
Rank Indicator Batch / Streaming Why streaming wins so much
1 msw 5490× Mesa Sine Wave requires FFT-style computation over the full window
2 cci 1393× Large rolling sum + mean deviation computed over every bar in batch
3 fisher 1348× Complex normalisation over entire input range
4 aroonosc 602× Full lookback window scan for highest/lowest on every batch call
5 ultosc 573× Three-period weighted TR average; all periods recomputed in batch
6 vhf 512× Rolling highest high / lowest low + sum of price changes
7 stoch 510× Rolling high/low window scan
8 willr 506× Rolling high/low window scan
9 aroon 473× Same as aroonosc
10 chandelierexit 454× Rolling highest high / lowest low + ATR across full window in batch