KunQuant

KunQuant is a optimizer, code generator and executor for financial expressions and factors, e.g. (close - open) /((high - low) + 0.001). The initial aim of it is to generate efficient implementation code for Alpha101 of WorldQuant and Alpha158 of Qlib. Some existing implementations of Alpha101 is straightforward but too simple. Hence we are developing KunQuant to provide optimizated code on a batch of general customized factors.

This project has mainly two parts: KunQuant and KunRunner. KunQuant is an optimizer & code generator written in Python. It takes a batch of financial expressions as the input and it generates highly optimized C++ code for computing these expressions. KunRunner is a supporting runtime library and Python wrapper to load and run the generated C++ code from KunQuant.

A typical workload of designing and running financial factors with KunQuant will be

1. Write the factors with KunQuant Python library
2. Use KunQuant to optimize the factors and transform them into C++ source code
3. Use cmake to compile the generated code
4. Load the genereted binary via KunRunner in Python code

Experiments show that KunQuant-generated code can be more than 170x faster than naive implementation based on Pandas. We ran Alpha001~Alpha101 with Pandas-based code and our optimized code. See results below:

Datatype	Pandas-based	KunQuant 1-thread	KunQuant 4-threads
Single precision (STs layout)	6.138s	0.083s	0.027s
Double precision (TS layout)	6.332s	0.120s	0.031s

The data was collected on 4-core Intel i7-7700HQ CPU, running synthetic data of 64 stocks with 260 rows of data in single precision float point data type. Environment:

OS=Ubuntu 22.04.3 on WSL2 on Windows 10
python=3.10.2
pandas=2.1.4
numpy=1.26.3
g++=11.4.0

Supported features of KunQuant

• Batch mode and stream mode for the input
• Double and single precision float point data type
• TS or STs memory layout as input/output in batch mode
• Python/C/C++ interfaces to call the factor computation functions

Why KunQuant is fast

• KunQuant parallelizes the computation for factors and uses SIMD (AVX2) to vectorize them.
• Redundant computation among factors are eliminated: Think what we can do with sum(x), avg(x), stddev(x)? The result of sum(x) is needed by all these factors. KunQuant also automatically finds if a internal result of a factor is used by other factors and try to reuse the results.
• Temp buffers are minimized by operator-fusion. For a factor like (a+b)/2, pandas and numpy will first compute the result of (a+b) and collect all the result in a buffer. Then, /2 opeator is applied on each element of the temp buffer of (a+b). This will result in large memory usage and bandwidth. KunQuant will generate C++ code to compute (a[i]+b[i])/2 in the same loop, to avoid the need to access and allocate temp memory.

Dependency

• pybind11 (automatically cloned via git as a submodule)
• Python (3.7+ with f-string and dataclass support)
• cmake
• A working C++ compiler with C++11 support (e.g. clang, g++, msvc)
• x86-64 CPU with at least AVX2-FMA instruction set
• Optionally requires AVX512 on CPU for better performance

Important node: For better performance compared with Pandas, KunQuant suggests to use a multiple of {blocking_len} as the number of stocks in inputs. For single-precision float type and AVX2 instruction set, blocking_len=8. That is, you are suggested to input 8, 16, 24, ..., etc. stocks in a batch, if your code is compiled with AVX2 (without AVX512) and float datatype. Other numbers of stocks are supported, with lower execution performance.

Compiling and running Alpha101

This section serves as am example for compiling an existing factor library: Alpha101 and running it. Building and running your own factors will be similar. If you are only interested in how you can run Alpha101 factors, this section is all you need. First, clone the KunQuant repo and make a new directory named build:

git clone https://github.com/Menooker/KunQuant --recursive
cd KunQuant
mkdir build
cd build

Then run cmake to configure the build:

cmake ..

If you want to use a non-default binary of Python executable, instead of the above command, run

cmake .. -DPYTHON_EXECUTABLE="PATH/TO/PYTHON/EXECUTABLE"

Build the code with cmake:

cmake --build . -- -j4

If the build is successful, you should be able to see in the terminal:

...
[100%] Built target Alpha101

You can find KunRunner.cpython-??-{x86_64-linux-gnu.so, amd64.pyd, darwin.so} and projects/{libAlpha101.so, libAlpha101.dylib} (on Linux/macOS) or projects/Release/Alpha101.dll (on Windows) in your build directory.

libAlpha101.so, Alpha101.dll or libAlpha101.dylib is the compiled code for Alpha101 factors on Linux, Windows or macOS. KunRunner is a Cpp extension for Python with helps to load the generated factor libraries. It also contains some supportive functions for the loaded libraries.

Before running Python, set the environment variable of PYTHONPATH:

On linux

export PYTHONPATH=$PYTHONPATH:/PATH/TO/KunQuant/build

On windows powershell

$env:PYTHONPATH+=";x:\PATH\TO\KunQuant\build\Release"

Note that /PATH/TO/KunQuant/build or x:\PATH\TO\KunQuant\build\Release should be the directory containing KunRunner.cpython-...{pyd,so}

Then in Python, import KunRunner and load the Alpha101 library:

import KunRunner as kr
lib = kr.Library.load("./projects/libAlpha101.so")
modu = lib.getModule("alpha_101")

Note that you need to give KunRunner a relative or absolute path of the factor library by replacing "./projects/libAlpha101.so" above.

Load your stock data. In this example, load from local pandas files. We assume the open, close, high, low, volumn and amount data for different stocks are stored in different files.

import pandas as pd

# we need a multiple of 8 number of stocks
watch_list = ["000002", "000063", ...]
num_stocks = len(watch_list)
assert(num_stocks % 8 == 0)
df = []

for stockid in watch_list:
    d = pd.read_hdf(f"{stockid}.hdf5")
    df.append(d)

print(df[0])

cols = df[0].columns.values
col2idx = dict(zip(cols, range(len(cols))))
print("columns to index", col2idx)
num_time = len(df[0])
print("dimension in time", num_time)

Here we printed the data frame of the first stock and the column-index mapping, it should look like:

                 open       high        low      close       volume        amount
date                                                                             
2020-01-02  32.799999  33.599998  32.509998  32.560001  101213040.0  3.342374e+09
2020-01-03  32.709999  32.810001  31.780001  32.049999   80553632.0  2.584310e+09
2020-01-06  31.750000  31.760000  31.250000  31.510000   87684056.0  2.761449e+09
...               ...        ...        ...        ...          ...           ...
2024-01-30  10.000000  10.050000   9.790000   9.790000   79792704.0  7.903654e+08
2024-01-31   9.770000   9.850000   9.560000   9.600000   67478864.0  6.527274e+08
2024-02-01   9.530000   9.660000   9.420000   9.440000   62786032.0  5.980486e+08

[993 rows x 6 columns]
columns to index {'open': 0, 'high': 1, 'low': 2, 'close': 3, 'volume': 4, 'amount': 5}
dimension in time 993

Transform your pandas data to numpy array of shape [features, stocks, time]. Feature here means the columns for open, close, high, low, volumn and amount.

import numpy as np

# [features, stocks, time]
collected = np.empty((len(col2idx), num_stocks, len(df[0])), dtype="float32")
for stockidx, data in enumerate(df):
    for colname, colidx in col2idx.items():
        mat = data[colname].to_numpy()
        collected[colidx, stockidx, :] = mat

Then an important step is to transpose the numpy array to shape [features, stocks//8, time, 8]. We split the axis of stocks into two axis [stocks//8, 8]. This step makes the memory layout of the numpy array match the SIMD length of AVX2, so that KunQuant can process the data in parallel in a single SIMD instruction. Notes:

• the number 8 here is the blocking_num of the compiled code. It is decided by the SIMD lanes of the data type and the instruction set (AVX2 or AVX512). By default, the example code of Alpha101 generates float dtype with AVX2. The register size of AVX2 is 256 bits, so the SIMD lanes of float should be 8.
• you can change the projects/Alpha101/generate.py to let the compiled code accept the simple matrix of [features, time, stocks] without the need of transposing in this step. See below [section](#Specifing Memory layouts and data types) for more details. Using TS layout may result slower execution of the factors.

# [features, stocks, time] => [features, stocks//8, 8, time] => [features, stocks//8, time, 8]
transposed = collected.reshape((collected.shape[0], -1, 8, collected.shape[2])).transpose((0, 1, 3, 2))
transposed = np.ascontiguousarray(transposed)

Now fill the input data in a dict

input_dict = dict()
for colname, colidx in col2idx.items():
    input_dict[colname] = transposed[colidx]

Create an executor and compute the factors!

# using 4 threads
executor = kr.createMultiThreadExecutor(4)
out = kr.runGraph(executor, modu, input_dict, 0, num_time)
print("Result of alpha101", out["alpha001"])
print("Shape of alpha101", out["alpha001"].shape)

Each output factors are computed in an array of shape [time, stocks]. The output of above code can be:

Result of alpha001 [[   nan    nan    nan ...    nan    nan    nan]
 [   nan    nan    nan ...    nan    nan    nan]
 [   nan    nan    nan ...    nan    nan    nan]
 ...
 [0.6875 0.1875 0.1875 ... 0.6875 0.6875 0.6875]
 [0.6875 0.1875 0.1875 ... 0.6875 0.6875 0.6875]
 [0.4375 1.     0.875  ... 0.4375 0.4375 0.4375]]
Shape of alpha001 (993, 8)

By default, runGraph will allocate an numpy array for each of the output factor. However, you can preallocate a numpy array and tell KunRunner to fill in this array instead of creating new ones.

outnames = modu.getOutputNames()
out_dict = dict()
# [Factors, Time, Stock]
sharedbuf = np.empty((len(outnames), num_time, num_stocks), dtype="float32")
for idx, name in enumerate(outnames):
    out_dict[name] = sharedbuf[idx]
out = kr.runGraph(executor, modu, input_dict, 0, num_time, out_dict)
# results are in "out" and "sharedbuf"

Note that the executors are reusable. A multithread executor is actually a thread pool inside. If you want to run on multiple batches of data, you don’t need to create new executors for each batch.

Customized factors

KunQuant is a tool for general expressions. You can further read Customize.md for how you can compile your own customized factors.

Streaming mode

KunQuant can be configured to generate factor libraries for streaming, when the data arrive one at a time. See Stream.md

Specifing Memory layouts and data types

The developers can choose the memory layout when compiling KunQuant factor libraries. The memory layout decribes how the input/output matrix is organized. Currently, KunQuant supports TS, STs and STREAM as the memory layout. In TS layout, the input and output data is in plain [num_time, num_stocks] 2D matrix. In STs with blocking_len = 8, the data should be transformed to [num_stocks//8, num_time, 8] for better performance. The STREAM layout is for the streaming mode. You can choose the input/output layout independently in compileit() function of generate.py, by the parameters compileit(..., input_layout="TS", output_layout="STs") for example. By default, the input layout is STs and the output layout is TS. For more info of customizing the factor compilation, see Customize.md.

KunQuant supports float and double data types. It can be selected by the dtype parameter of compileit() in your own generate.py.

If CMake Option -DKUN_AVX512 is ON (by default is OFF), the blocking_len for dtype='float' can be 8 or 16, and for dtype='double' can be 4 or 8. If -DKUN_AVX512 is not specified or is OFF, the blocking_len for dtype='float' should only be 8, and for dtype='double' should be 4.

Enabling AVX512

This project by default turns off AVX512, since this intruction set is not yet well adopted. If you are sure your CPU has AVX512, you can turn it on by adding cmake option -DKUN_AVX512=ON when running cmake command above. This will enable AVX512 features when compiling the KunQuant generated code. Some speed-up over AVX2 mode are expected.

In your customized project, you need to specify blocking_len parameter of in compileit() function of generate.py to enable AVX512. See above [section](#Specifing Memory layouts and data types). The example projects Alpha101, Alpha101Stream, Alpha158 in projects/ will detect if -DKUN_AVX512=ON and automatically set blocking_len to use AVX2 or AVX512. Please note that blocking_len will affect the STs format.

There are some other CPU instruction sets that is optional for KunQuant. You can turn on AVX512DQ and AVX512VL to accelerate some parts of KunQuant-generated code. To enable them, add -DKUN_AVX512DQ=ON and -DKUN_AVX512VL=ON in cmake options respectively.

To see if your CPU supports AVX512 (and AVX512DQ and AVX512VL), you can run command lscpu in Linux and check the outputs.

Enabling AVX512 will slightly improve the performance, if it is supported by the CPU. Experiments only shows ~1% performance gain for 16-threads of AVX512 on Icelake, testing on double-precision Alpha101, with 128 stocks and time length of 12000. A single thread running the same task shows 5% performance gain on AVX512.

Operator definitions

See Operators.md

Testing and validation

Unit tests for some of the internal IR transformations:

python tests/test.py
python tests/test2.py

Unit tests for C++ runtime:

python tests/test_runtime.py

To run the runtime UTs, you need to make sure you have built the cmake target KunTest by

cmake --build . --target KunTest

Correctness test of Alpha101

# current dir should be at the base directory of KunQuant
python tests/test_alpha101.py

The input data are randomly genereted data and the results are checked against a modified (corrected) version of Pandas-based code. Note that some of the factors like alpha013 are very sensitive to numerical changes in the intermeidate results, because rank operators are used. The result may be very different after rank even if the input is very close. Hence, the tolerance of these factors will be high to avoid false positives.

To test Alpha158, you need first download the input data and reference result files: alpha158.npz and input.npz.

Then run

# current dir should be at the base directory of KunQuant
python tests/test_alpha158.py --inputs /PATH/TO/input.npz --ref /PATH/TO/alpha158.npz 

This script runs alpha158 with double precision mode in KunQuant. It feeds the library with predefined values from input.npz and check against the result with alpha158.npz, which is computed by qlib.

To generate another Alpha158 result with another randomly generated input, you can run

# current dir should be at the base directory of KunQuant
python ./tests/gen_alpha158.py --tmp /tmp/a158 --qlib /path/to/source/of/qlib --out /tmp

It will create the random input at /tmp/input.npz and result at /tmp/alpha158.npz

Using C-style APIs

KunQuant provides C-style APIs to call the generated factor code in shared libraries. See CAPI.md

Acknowledgement

The implementation and testing code for Alpha101 is based on https://github.com/yli188/WorldQuant_alpha101_code

The implementation code for Alpha158 is based on https://github.com/microsoft/qlib/blob/main/qlib/contrib/data/handler.py. Licensed under the MIT License.

The AVX vector operators at cpp/KunSIMD/cpu was developed based on x86simd as a component of GraphCompiler, a backend of oneDNN Graph API. Licensed under the Apache License, Version 2.0 (the "License").