function CDLDOJISTAR works ok?

[phoenixxz]

hi, I tested the function CDLDOJISTAR to find '+' shaped k lines, I found that not all the '+' shaped k lines were selected. Some "+" shaped k lines were lost? Who has met the same problem? Or it's the problem
of my codes?

import tushare as ts
import talib as ta
import numpy as np

def back_track(code):
data = ts.get_h_data(code,start='2017-05-01')
high = np.array(data['high'], dtype=float)
low = np.array(data['low'], dtype=float)
open = np.array(data['open'], dtype=float)
close = np.array(data['close'], dtype=float)
date = data.index
integer = ta.CDLDOJISTAR(open, high, low, close)
for j in range(0, len(integer), 1):
if integer[j] != 0:
print str(integer[j]) + "," + code + "," + str(date[j]) + "," + str(j) +'\r\n'

back_track('000728')

[mrjbq7]

You can look at the source code and see if it does what you expect:

https://sourceforge.net/p/ta-lib/code/HEAD/tree/trunk/ta-lib/c/src/ta_func/ta_CDLDOJISTAR.c

[bkcollection]

@phoenixxz please refer to my thread
#119

[mario4tier]

Sorry for the very long delay on this one. Two notes before the answer:

This analysis is AI-assisted. It rests on the mechanism plus substitute datasets, because the data behind your run isn't retrievable — ts.get_h_data was removed from tushare years ago, so code 000728 from 2017-05-01 can't be pulled today. Attaching the input data to a report is genuinely helpful; it's usually the difference between "here is exactly what happened on your bars" and "here is why this happens in general".

Short version: not a bug. CDLDOJISTAR is not a doji finder — the "+" shaped candles you're looking for are CDLDOJI.

Why some "+" candles are skipped

CDLDOJISTAR is a two-candle reversal pattern. A doji alone is not enough; it needs all three:

1. the previous candle has a long real body
2. the current candle is a doji
3. the doji's real body gaps away from the previous body — up if the previous candle was white, down if it was black

Your loop asks for pattern 2 only, so everything failing 1 or 3 looks "lost".

Measured over three real daily series (~257 bars each, from #119's attachments):

	CDLDOJI	CDLDOJISTAR
KESM	34	2
MFLOUR	73	4
PBBANK	33	3
total	140	9

94% of doji candles are correctly not doji stars. Of the 131 skipped: 31 failed the long-previous-body test, 39 the gap test, 61 both.

A concrete example — KESM 2015-10-30, a perfect "+" with open == close:

prev O=4.99 C=5.05 (white, body 0.06)
this O=5.05 H=5.05 L=4.92 C=5.05 CDLDOJI = 100 CDLDOJISTAR = 0

The doji opens exactly at the previous close, so there is no real-body gap. Correctly rejected. Compare the two bars that do qualify:

2016-03-21 prev O=4.42 C=4.33 (black) -> doji at 4.29, gaps down = +100
2016-08-17 prev O=7.71 C=7.98 (white) -> doji at 8.00, gaps up = -100

What to call instead

• CDLDOJI — any "+" shaped candle. This is what you want.
• CDLLONGLEGGEDDOJI, CDLDRAGONFLYDOJI, CDLGRAVESTONEDOJI, CDLRICKSHAWMAN — specific doji shapes.
• CDLDOJISTAR, CDLMORNINGDOJISTAR, CDLEVENINGDOJISTAR, CDLHARAMICROSS, CDLTRISTAR — multi-candle patterns that contain a doji.

One related knob: what counts as a doji is itself configurable. BodyDoji defaults to "real body no larger than 10% of the 10-bar average high-low range", so a near-"+" with a small non-zero body on a wide-range day can also fall outside it.

On the #119 cross-reference

@bkcollection pointed you at #119, but that is a separate and unrelated scenario. That report turned out to be a one-position column shift in that reporter's own aggregation script — df_HAMMER was being filled with CDLGRAVESTONEDOJI's result, and so on down the list. Nothing to do with what you are seeing here, which is simply CDLDOJISTAR doing what a doji star is defined to do. #119 is answered separately.

No defect found in CDLDOJISTAR.