Quote from abattia:
I need to store and work with large numbers of ticks (last traded price) in memory. Can I store these ticks as float rather than double?
I can't think of any tick requiring greater than 7 digits of precision. Are there any?
When I use the ticks for calculations I will convert to double precision, but will keep them stored in memory as float. Am I missing anything?
Quote from hftvol:
Ok, care to show how its done correctly with your approach?
Lets say you have a loop and divide within the loop by 1000.123456789. You not only need to carry around the multiplier/char (which indicated how to convert back and forth between your int and double/float) but you have to do so 10 times and hold potentially 10 different such multipliers or at least have to make 10 separate adjustments to such factor.
So, it really depends, but for most all financial applications floating point representation is way precise enough, pending choice of correct variable type (double vs. float, ...).
Your approach wastes not only memory but also computational power (imagine you do that in a loop that iterates over tick based data of the magnitude of million observations and that not just for one price point but bid/ask/last/... your code will run several orders of magnitude slower than just using the correct floating point type which is way precise enough for most all calculations pertaining to finance.
Quote from hft_boy:
Luckily in finance they used fixed point (dollars, cents and bps), so you can map the numbers used directly to integers. In fact, a lot of exchanges (e.g. Nasdaq, NYSE) distribute data where the prices are given in a fixed point representation rather than floating point, just so nobody gets the wrong idea.
And, the multiplier is almost always the same (100 or 10000, depending on the market / exchange), so you don't need to carry that around but just keep it as a compile time constant. Also, converting once back and forth (when you read the data and when you need to output something) is quite a cheap operation, especially when compared to how expensive floating point operations are. In fact, if you're doing a lot of floating point operations, it will probably actually be faster to use integers internally.
And sure, adding two floating point numbers and dividing by another is going to be accurate within like 2 ulp. But given that simple things like computing a moving average are numerically unstable (granted usually precise with 52-bits in 8-byte doubles), there are a lot of advantages to using fixed point.
Lastly, keep in mind also that 7 digits of precision in decimal is about 23 digits of precision in binary, which is right on the edge of exceeding the precision of 32-bit floats, which the OP was planning on using.
Quote from hftvol:
this is not true. There are not A LOT of exchanges that distribute fixed points, there are some but not many. And its also not true that the multiplier is almost always the same, go to Nanex (which is not an exchange but a data provider who uses fixed point data distribution and conversions), they use over 13 (or was it 19) different multipliers for base 10 and the same number for base 2, LOL.
Quote from hftvol:
this is not true. There are not A LOT of exchanges that distribute fixed points, there are some but not many. And its also not true that the multiplier is almost always the same, go to Nanex (which is not an exchange but a data provider who uses fixed point data distribution and conversions), they use over 13 (or was it 19) different multipliers for base 10 and the same number for base 2, LOL.
Quote from hftvol:
Bloomberg P-Pipe and Reuters RMDS both do disseminate their quotes as floating points. Both are the market leaders in consolidated data feeds that connect to over 200 exchanges, both are heavily used by hedge funds, buy side firms and investment banks. Most direct exchange feeds also use floating point price data. Saying all use integer type feeds is factually wrong.
Quote from Makis:
I have been working with market data feeds for a living -both direct and vendor feeds- for well over 18 years now. I did have stints with Bloomberg and Reuters as well. As a matter of fact I have worked FOR b-pipe. For what is worth, let me assure you that they disseminate prices as integers. It is their API on the client side that converts integers to doubles.
My question was if you know another vendor, other than nanex, that can process an entire feed on a single computer. You don't claim that b-pipe or RMDS can do that, right?
I pretty much know most major exchange feed specs by heart. None disseminates prices as doubles (again with exception any low budget ascii feeds that may exist for low budget retail users)
Please show me where I am factually wrong.
Can you name an exchange feed that disseminates prices as floating points?
Bellow are some spec links I have handy so you do not waste your time looking into those.
CTS (NYSE TRADES) http://www.nyxdata.com/nysedata/Def...ule=6672&EntryId=188107&Command=Core_Download
Page 118, section 11.84 TRADE PRICE
CQS (NYSE QUOTES) http://www.nyxdata.com/nysedata/Def...ule=6672&EntryId=188106&Command=Core_Download
Page 44, section 11.8 BEST OFFER PRICE
ITCH (NASDAQ) http://www.nasdaqtrader.com/content/technicalsupport/specifications/dataproducts/NQTV-ITCH-V4_1.pdf
page 2, Section 3 Data Types
"Prices are integer fields. When converted to a decimal format, prices are in fixed point format with 6
whole number places followed by 4 decimal digits. The maximum price in TotalView-ITCH is
200,000.0000 (decimal, 77359400 hex)."
BATS http://cdn.batstrading.com/resources/membership/BATS_MC_PITCH_Specification.pdf
page 10, section 2.2
"Binary Long Price fields are unsigned Little Endian encoded 8 byte binary fields with 4
implied decimal places (denominator = 10,000)."
OPRA http://www.opradata.com/specs/OPRA_Binary_Part_Spec_1.2_032013.pdf
Page 42, section 7.10 Last Price
"4 bytes, signed integer. The Last Price is the whole and decimal portion of the Last Price information with the
Premium Price Denominator Code determining the location of the decimal point."