Decompress and read Dukascopy .bi5 tick files
Question:
I need to open a .bi5 file and read the contents to cut a long story short. The problem: I have tens of thousands of .bi5 files containing time-series data that I need to decompress and process (read, dump into pandas).
I ended up installing Python 3 (I use 2.7 normally) specifically for the lzma library, as I ran into compiling nightmares using the lzma back-ports for Python 2.7, so I conceded and ran with Python 3, but with no success. The problems are too numerous to divulge, no one reads long questions!
I have included one of the .bi5 files, if someone could manage to get it into a Pandas Dataframe and show me how they did it, that would be ideal.
ps the fie is only a few kb, it will download in a second. Thanks very much in advance.
(The file)
http://www.filedropper.com/13hticks
Answers:
The code below should do the trick. First, it opens a file and decodes it in lzma and then uses struct to unpack the binary data.
import lzma
import struct
import pandas as pd
def bi5_to_df(filename, fmt):
chunk_size = struct.calcsize(fmt)
data = []
with lzma.open(filename) as f:
while True:
chunk = f.read(chunk_size)
if chunk:
data.append(struct.unpack(fmt, chunk))
else:
break
df = pd.DataFrame(data)
return df
The most important thing is to know the right format. I googled around and tried to guess and '>3i2f' (or >3I2f) works quite good. (It’s big endian 3 ints 2 floats. What you suggest: 'i4f' doesn’t produce sensible floats – regardless whether big or little endian.) For struct and format syntax see the docs.
df = bi5_to_df('13h_ticks.bi5', '>3i2f')
df.head()
Out[177]:
0 1 2 3 4
0 210 110218 110216 1.87 1.12
1 362 110219 110216 1.00 5.85
2 875 110220 110217 1.00 1.12
3 1408 110220 110218 1.50 1.00
4 1884 110221 110219 3.94 1.00
Update
To compare the output of bi5_to_df with https://github.com/ninety47/dukascopy,
I compiled and run test_read_bi5 from there. The first lines of the output are:
time, bid, bid_vol, ask, ask_vol
2012-Dec-03 01:00:03.581000, 131.945, 1.5, 131.966, 1.5
2012-Dec-03 01:00:05.142000, 131.943, 1.5, 131.964, 1.5
2012-Dec-03 01:00:05.202000, 131.943, 1.5, 131.964, 2.25
2012-Dec-03 01:00:05.321000, 131.944, 1.5, 131.964, 1.5
2012-Dec-03 01:00:05.441000, 131.944, 1.5, 131.964, 1.5
And bi5_to_df on the same input file gives:
bi5_to_df('01h_ticks.bi5', '>3I2f').head()
Out[295]:
0 1 2 3 4
0 3581 131966 131945 1.50 1.5
1 5142 131964 131943 1.50 1.5
2 5202 131964 131943 2.25 1.5
3 5321 131964 131944 1.50 1.5
4 5441 131964 131944 1.50 1.5
So everything seems to be fine (ninety47’s code reorders columns).
Also, it’s probably more accurate to use '>3I2f' instead of '>3i2f' (i.e. unsigned int instead of int).
Did you try using numpy as to parse the data before transfer it to pandas. Maybe is a long way solution, but I will allow you to manipulate and clean the data before you made the analysis in Panda, also the integration between them are pretty straight forward,
import requests
import struct
from lzma import LZMADecompressor, FORMAT_AUTO
# for download compressed EURUSD 2020/06/15/10h_ticks.bi5 file
res = requests.get("https://www.dukascopy.com/datafeed/EURUSD/2020/06/15/10h_ticks.bi5", stream=True)
print(res.headers.get('content-type'))
rawdata = res.content
decomp = LZMADecompressor(FORMAT_AUTO, None, None)
decompresseddata = decomp.decompress(rawdata)
firstrow = struct.unpack('!IIIff', decompresseddata[0: 20])
print("firstrow:", firstrow)
# firstrow: (436, 114271, 114268, 0.9399999976158142, 0.75)
# time = 2020/06/15/10h + (1 month) + 436 milisecond
secondrow = struct.unpack('!IIIff', decompresseddata[20: 40])
print("secondrow:", secondrow)
# secondrow: (537, 114271, 114267, 4.309999942779541, 2.25)
# time = 2020/06/15/10h + (1 month) + 537 milisecond
# ask = 114271 / 100000 = 1.14271
# bid = 114267 / 100000 = 1.14267
# askvolume = 4.31
# bidvolume = 2.25
# note that 00 -> is january
# "https://www.dukascopy.com/datafeed/EURUSD/2020/00/15/10h_ticks.bi5" for january
# "https://www.dukascopy.com/datafeed/EURUSD/2020/01/15/10h_ticks.bi5" for february
# iterating
print(len(decompresseddata), int(len(decompresseddata) / 20))
for i in range(0, int(len(decompresseddata) / 20)):
print(struct.unpack('!IIIff', decompresseddata[i * 20: (i + 1) * 20]))
In case someone used endpoint
https://datafeed.dukascopy.com/datafeed/EURUSD/2022/11/06/BID_candles_min_1.bi5
the format is ‘>IIIIIf’ (big endian, 5x integer, 1x float )
and the columns are Seconds, O, H, L, C, V
I need to open a .bi5 file and read the contents to cut a long story short. The problem: I have tens of thousands of .bi5 files containing time-series data that I need to decompress and process (read, dump into pandas).
I ended up installing Python 3 (I use 2.7 normally) specifically for the lzma library, as I ran into compiling nightmares using the lzma back-ports for Python 2.7, so I conceded and ran with Python 3, but with no success. The problems are too numerous to divulge, no one reads long questions!
I have included one of the .bi5 files, if someone could manage to get it into a Pandas Dataframe and show me how they did it, that would be ideal.
ps the fie is only a few kb, it will download in a second. Thanks very much in advance.
(The file)
http://www.filedropper.com/13hticks
The code below should do the trick. First, it opens a file and decodes it in lzma and then uses struct to unpack the binary data.
import lzma
import struct
import pandas as pd
def bi5_to_df(filename, fmt):
chunk_size = struct.calcsize(fmt)
data = []
with lzma.open(filename) as f:
while True:
chunk = f.read(chunk_size)
if chunk:
data.append(struct.unpack(fmt, chunk))
else:
break
df = pd.DataFrame(data)
return df
The most important thing is to know the right format. I googled around and tried to guess and '>3i2f' (or >3I2f) works quite good. (It’s big endian 3 ints 2 floats. What you suggest: 'i4f' doesn’t produce sensible floats – regardless whether big or little endian.) For struct and format syntax see the docs.
df = bi5_to_df('13h_ticks.bi5', '>3i2f')
df.head()
Out[177]:
0 1 2 3 4
0 210 110218 110216 1.87 1.12
1 362 110219 110216 1.00 5.85
2 875 110220 110217 1.00 1.12
3 1408 110220 110218 1.50 1.00
4 1884 110221 110219 3.94 1.00
Update
To compare the output of bi5_to_df with https://github.com/ninety47/dukascopy,
I compiled and run test_read_bi5 from there. The first lines of the output are:
time, bid, bid_vol, ask, ask_vol
2012-Dec-03 01:00:03.581000, 131.945, 1.5, 131.966, 1.5
2012-Dec-03 01:00:05.142000, 131.943, 1.5, 131.964, 1.5
2012-Dec-03 01:00:05.202000, 131.943, 1.5, 131.964, 2.25
2012-Dec-03 01:00:05.321000, 131.944, 1.5, 131.964, 1.5
2012-Dec-03 01:00:05.441000, 131.944, 1.5, 131.964, 1.5
And bi5_to_df on the same input file gives:
bi5_to_df('01h_ticks.bi5', '>3I2f').head()
Out[295]:
0 1 2 3 4
0 3581 131966 131945 1.50 1.5
1 5142 131964 131943 1.50 1.5
2 5202 131964 131943 2.25 1.5
3 5321 131964 131944 1.50 1.5
4 5441 131964 131944 1.50 1.5
So everything seems to be fine (ninety47’s code reorders columns).
Also, it’s probably more accurate to use '>3I2f' instead of '>3i2f' (i.e. unsigned int instead of int).
Did you try using numpy as to parse the data before transfer it to pandas. Maybe is a long way solution, but I will allow you to manipulate and clean the data before you made the analysis in Panda, also the integration between them are pretty straight forward,
import requests
import struct
from lzma import LZMADecompressor, FORMAT_AUTO
# for download compressed EURUSD 2020/06/15/10h_ticks.bi5 file
res = requests.get("https://www.dukascopy.com/datafeed/EURUSD/2020/06/15/10h_ticks.bi5", stream=True)
print(res.headers.get('content-type'))
rawdata = res.content
decomp = LZMADecompressor(FORMAT_AUTO, None, None)
decompresseddata = decomp.decompress(rawdata)
firstrow = struct.unpack('!IIIff', decompresseddata[0: 20])
print("firstrow:", firstrow)
# firstrow: (436, 114271, 114268, 0.9399999976158142, 0.75)
# time = 2020/06/15/10h + (1 month) + 436 milisecond
secondrow = struct.unpack('!IIIff', decompresseddata[20: 40])
print("secondrow:", secondrow)
# secondrow: (537, 114271, 114267, 4.309999942779541, 2.25)
# time = 2020/06/15/10h + (1 month) + 537 milisecond
# ask = 114271 / 100000 = 1.14271
# bid = 114267 / 100000 = 1.14267
# askvolume = 4.31
# bidvolume = 2.25
# note that 00 -> is january
# "https://www.dukascopy.com/datafeed/EURUSD/2020/00/15/10h_ticks.bi5" for january
# "https://www.dukascopy.com/datafeed/EURUSD/2020/01/15/10h_ticks.bi5" for february
# iterating
print(len(decompresseddata), int(len(decompresseddata) / 20))
for i in range(0, int(len(decompresseddata) / 20)):
print(struct.unpack('!IIIff', decompresseddata[i * 20: (i + 1) * 20]))
In case someone used endpoint
https://datafeed.dukascopy.com/datafeed/EURUSD/2022/11/06/BID_candles_min_1.bi5
the format is ‘>IIIIIf’ (big endian, 5x integer, 1x float )
and the columns are Seconds, O, H, L, C, V