chronos2-cpp

End-to-end C++ pipeline for realised-volatility forecasting: Bloomberg BLPAPI ingest → SQLite → feature engineering → Chronos-2 inference via ONNX Runtime, benchmarked against a from-scratch TCN, HAR-RV and persistence with Diebold-Mariano tests.

Installation

The project is currently configured for Windows using MSYS2 UCRT64, GCC/G++, CMake, Ninja, Eigen3, SQLite, ONNX Runtime, and the Bloomberg BLPAPI C++ SDK.

1. Clone the repository

git clone https://github.com/h3dk4ndi/chronos2-cpp.git
cd chronos2-cpp

Or manually:

• Open the repository in GitHub.
• Click Code.
• Select Download ZIP.
• Extract the ZIP archive.
• Open the extracted chronos2-cpp folder in VS Code.

2. Install MSYS2 UCRT64 dependencies

Open an MSYS2 UCRT64 terminal; update the MSYS2 environment:

pacman -Syu

If MSYS2 asks you to close the terminal while updating core packages, close it, reopen the UCRT64 terminal, and run again:

pacman -Syu

You ideally want something like:

:: Synchronizing package databases...
ucrt64 is up to date
mingw64 is up to date
clang64 is up to date
msys is up to date

:: Starting core system upgrade...
 there is nothing to do
:: Starting full system upgrade...
 there is nothing to do

Install the required packages:

pacman -S mingw-w64-ucrt-x86_64-cmake \
          mingw-w64-ucrt-x86_64-ninja \
          mingw-w64-ucrt-x86_64-eigen3 \
          mingw-w64-ucrt-x86_64-sqlite3 \
          mingw-w64-ucrt-x86_64-onnxruntime

Verify the development environment:

gcc --version
g++ --version
cmake --version
ninja --version

You should see version information similar to:

gcc.exe (Rev8, Built by MSYS2 project) 16.x.x
g++.exe (Rev8, Built by MSYS2 project) 16.x.x
cmake version 4.x.x
1.x.x

3. Verify ONNX Runtime

The ONNX Runtime C++ headers should be installed under the UCRT64 environment. Run:

find /ucrt64 -name "onnxruntime_cxx_api.h"

A correct installation should return: /ucrt64/include/onnxruntime/onnxruntime_cxx_api.h The corresponding runtime DLLs should also exist:

/ucrt64/bin/onnxruntime.dll
/ucrt64/bin/onnxruntime_providers_shared.dll

4. Download the Bloomberg BLPAPI C++ SDK

Download the Bloomberg BLPAPI C++ SDK for Windows: Bloomberg SDK C++ for Windows

Extract the archive. For version 3.26.7.1, the resulting SDK directory should have approximately the following structure:

blpapi_cpp_3.26.7.1/
├── include/
│   ├── blpapi_session.h
│   └── ...
└── lib/
    ├── blpapi3_64.dll
    ├── blpapi3_64.lib
    ├── blpapi3_32.dll
    └── blpapi3_32.lib

The Bloomberg SDK does not need to be installed system-wide or copied into the repository. The build system only needs the location of the extracted SDK.

The BLPAPI SDK provides the API libraries only. Bloomberg market-data retrieval still requires access to Bloomberg Professional / Bloomberg Terminal with an active Bloomberg session.

5. Configure local paths in build.bat

Open build.bat and set the paths to your local UCRT64 and Bloomberg SDK installations. For example:

set "UCRT64=C:\path\to\msys64\ucrt64"

set "BLPAPI_ROOT=C:\Users\yourName\Downloads\blpapi_cpp_3.26.7.1-windows\blpapi_cpp_3.26.7.1"

BLPAPI_ROOT must point to the directory that directly contains:

include\
lib\

6. Configure VS Code IntelliSense

In VS Code press: Ctrl + Shift + P and select: C/C++: Edit Configurations (JSON)

Configure .vscode/c_cpp_properties.json using the actual location of your UCRT64 installation:

{
    "configurations": [
        {
            "name": "UCRT64",
            "compilerPath": "C:/path/to/msys64/ucrt64/bin/g++.exe",
            "includePath": [
                "${workspaceFolder}/**",
                "C:/path/to/msys64/ucrt64/include",
                "C:/path/to/msys64/ucrt64/include/onnxruntime"
            ],
            "cppStandard": "c++17",
            "cStandard": "c17",
            "intelliSenseMode": "windows-gcc-x64"
        }
    ],
    "version": 4
}

If VS Code asks which compiler should be used for IntelliSense, select the UCRT64 g++.exe installation.

If stale include errors remain:

Ctrl + Shift + P
→ C/C++: Reset IntelliSense Database

and then reload the VS Code window.

Additionally, configure the default C++ compiler used by the VS Code C/C++ extension (VS Code's settings.json, not c_cpp_properties.json). Open:

Ctrl + Shift + P
→ Preferences: Open User Settings (JSON)

and add:

{
    "C_Cpp.default.compilerPath": "C:/path/to/msys64/ucrt64/bin/g++.exe"
}

7. Export the Chronos-2 ONNX model

The ONNX model weights are not stored directly in the repository. Run:

models/onnx_chronos_v2.ipynb

to export the Chronos-2 model.

After export, the models directory must contain:

models/
├── chronos2.onnx
└── chronos2.onnx.data

Both files are required. chronos2.onnx contains the ONNX computational graph, while chronos2.onnx.data contains the external model parameters used by ONNX Runtime.

If chronos2.onnx.data is absent, ONNX Runtime will fail with an external-data-path error.

8. Build and run

Open PowerShell in the repository root and run:

.\build.bat

The build script performs:

CMake configuration
        ↓
Ninja compilation
        ↓
Bloomberg DLL deployment
        ↓
ONNX Runtime DLL deployment
        ↓
chronos2.exe

The generated executable is located at:

build\chronos2.exe

A successful run should produce output similar to:

XAU Curncy         7806 rows
XAG Curncy         7804 rows
...
TOTAL             69465

[split] train [...] purged 21 test [...]
[chronos2] quantiles=21 horizon=21 median_idx=10

Appendix: build.bat

The repository uses a small Windows batch script to configure, compile, deploy the required runtime libraries, and execute the project.

Before use, update the UCRT64 and BLPAPI_ROOT variables to match the local installation paths.

@echo off
setlocal

cd /d "%~dp0"

set "UCRT64=C:\Users\userName\Downloads\c inst\ucrt64"

set "BLPAPI_ROOT=C:\Users\userName\Downloads\blpapi_cpp_3.26.7.1-windows\blpapi_cpp_3.26.7.1"

set "PATH=%UCRT64%\bin;%PATH%"

if exist build (
    rmdir /s /q build
)

echo === Paths ===
echo UCRT64      = %UCRT64%
echo BLPAPI_ROOT = %BLPAPI_ROOT%
echo.

echo === Configuring ===

cmake -S . -B build ^
    -G Ninja ^
    -DCMAKE_BUILD_TYPE=Release ^
    -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ^
    -DUSE_BLPAPI=ON ^
    "-DONNXRUNTIME_ROOT=%UCRT64%" ^
    "-DBLPAPI_ROOT=%BLPAPI_ROOT%"

if errorlevel 1 goto :error

echo.
echo === Building ===

cmake --build build

if errorlevel 1 goto :error

echo.
echo === Copying runtime DLLs ===

copy /Y "%BLPAPI_ROOT%\lib\blpapi3_64.dll" "build\blpapi3_64.dll" >nul
copy /Y "%UCRT64%\bin\onnxruntime.dll" "build\onnxruntime.dll" >nul
copy /Y "%UCRT64%\bin\onnxruntime_providers_shared.dll" "build\onnxruntime_providers_shared.dll" >nul

echo.
echo === Running chronos2 ===
echo.

build\chronos2.exe

if errorlevel 1 goto :runerror

echo.
echo === Finished successfully ===

pause
exit /b 0

:error
echo.
echo === BUILD FAILED ===
pause
exit /b 1

:runerror
echo.
echo === PROGRAM EXITED WITH AN ERROR ===
pause
exit /b 1

Contributors

Special thanks to @ByteJoseph for their contributions, feedback, and support during the development of this project.

Additional contributors and their specific contributions will be acknowledged here as the project evolves.

Layout

Path	Contents
include/config.hpp	ROLL_W, CONTEXT, TEST_FRAC
include/types.hpp	InstrumentMeta
include/sqlite_storage.hpp	SQLite — blp_data, instrument_meta, prep_data
include/bloomberg_client.hpp	Bloomberg — historical + reference data requests
include/stationarity.hpp	Adfuller (Eigen), FracDiff (de Prado)
include/split.hpp	purged chronological train/val/test split
include/preprocessing.hpp	rolling primitives + RV estimators, semivariance, jumps, leverage
include/chronos2_onnx.hpp	Chronos2ONNX — ORT session wrapper
include/study.hpp	context matrix assembly and windowing
include/evaluation.hpp	QLIKE, HAR-RV, Diebold-Mariano
tcnvol/	NumPy TCN — layers, weight-norm dilated causal convs, AdamW, trainer
run_tcn.py	TCN entry point; writes per-origin QLIKE to tcn_loss.csv

Features

BuildMatrix() assembles a 14-row context matrix. Row 0 is the target series; rows 1-13 are covariates supplied to Chronos-2 through group attention and to the TCN as input channels.

Rows	Contents	Transform
0	close-to-close RV (target)	log
1-4	Parkinson, Garman-Klass, Rogers-Satchell, Yang-Zhang	log
5	returns	raw
6-7	negative / positive realised semivariance	log
8	bipower variation	log
9-13	signed jump, leverage, 5d leverage mean, jump component, relative jump	raw

Rows 9-13 can be zero or negative, so they are passed unlogged.

Requirements

• C++17
• Eigen 3
• SQLite 3
• ONNX Runtime 1.26+
• Bloomberg BLPAPI
• Python 3 + NumPy (TCN only)

Model weights

models/chronos2.onnx ships in the repo. The external weights file (~478 MB for the 120M base) exceeds GitHub's file limit and is not committed — download it separately and place it beside the graph as models/chronos2.onnx.data. The filename is recorded inside the graph; the ORT session constructor throws if it does not match exactly.

models/onnx_chronos_v2.ipynb regenerates the graph. Point MODEL at a local folder to export fine-tuned weights instead of the base checkpoint; nothing else in the notebook changes.

Run

./chronos2.exe        # Chronos-2, HAR-RV, persistence walk-forward
python run_tcn.py     # TCN — trains, evaluates, writes tcn_loss.csv

Both read prep_data from blp.db and score the same 2,321 origins.

Results

The final experiment uses XAU/USD only as the forecasting asset. Although additional Bloomberg series were collected during data preparation, the forecasting models use the 14 engineered features derived from XAU/USD rather than cross-asset inputs.

The final evaluation uses a purged train/test split with a 21-day forecast horizon:

[split] train [0,5443)  purged 21  test [5464,7806)
[test origins] 2321

Forecast performance:

Model	QLIKE
TCN	0.2441
HAR-RV	0.2760
Chronos-2, last-of-path	0.2787
Chronos-2, mean-of-path	0.2912
Persistence	0.3308

The TCN achieved the lowest QLIKE, improving on HAR-RV by approximately 11.5%.

Diebold-Mariano tests did not show a statistically significant difference between TCN and HAR-RV at the 5% level ($p = 0.1397$). TCN nevertheless significantly outperformed persistence ($p = 0.0422$).

Chronos-2 substantially outperformed persistence when using the mean forecast path ($p < 0.001$), but did not significantly outperform HAR-RV.

The TCN was trained using:

features = 14
window   = 512
epochs   = 50
batch    = 64
patience = 10

Early stopping selected the model at the minimum validation QLIKE:

best validation QLIKE = 0.1256
test QLIKE            = 0.2441