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.
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.
| 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 |
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.
- C++17
- Eigen 3
- SQLite 3
- ONNX Runtime 1.26+
- Bloomberg BLPAPI
- Python 3 + NumPy (TCN only)
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.
./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.
The database contains 69,465 daily observations across nine Bloomberg instruments. Evaluation was conducted on XAU Curncy, using 7,806 observations and a purged chronological test set of 2,321 forecast origins. Each model forecasts 21-day realised volatility; lower QLIKE indicates better predictive accuracy.
| Model | QLIKE | Skill vs HAR-RV |
|---|---|---|
| TCN (trained from scratch) | 0.2441 | +11.54% |
| HAR-RV | 0.2760 | — |
| Chronos-2 zero-shot (last of path) | 0.2787 | −0.99% |
| Chronos-2 zero-shot (mean of path) | 0.2912 | −5.52% |
| Persistence | 0.3308 | −19.87% |
The TCN achieved the lowest QLIKE, improving upon HAR-RV by 11.54%. However, this improvement was not statistically significant under the Diebold–Mariano test (DM = −1.48, p = 0.140). Its advantage over zero-shot Chronos-2 was likewise not significant (p = 0.235), although it significantly outperformed persistence at the 5% level (p = 0.042).
Zero-shot Chronos-2 using the terminal forecast was statistically indistinguishable from HAR-RV (DM = 0.13, p = 0.896), despite receiving no task-specific training. The mean-of-path Chronos-2 forecast also significantly outperformed persistence (DM = −4.04, p < 0.001). Overall, the TCN produced the strongest point estimate, while the statistical tests did not establish a significant difference in predictive accuracy among the TCN, HAR-RV and zero-shot Chronos-2.
Zero-shot Chronos-2 is statistically indistinguishable from HAR-RV (DM = 0.13, p = 0.90) despite never having seen gold. It beats persistence decisively.
Fine-tuning was attempted with LoRA on the attention projections (2.4M of 122M parameters trainable) over 4,911 purged training windows. The objective was pinball loss averaged across all 21 quantiles and all 21 horizon steps — the pretraining objective, not the evaluation target. Path-mean accuracy improved 3.4%; terminal-step accuracy, which is what QLIKE measures here, degraded 21%. The training objective has to match the evaluation target.
The TCN is a 7-block weight-normalised dilated causal network, dilations
[1,2,4,8,16,32,64] giving a receptive field of 509, roughly 8.6k parameters
against Chronos-2's 121.8M. It trains from scratch on the same 4,326 windows
that were insufficient to adapt the foundation model.
- The TCN result is a single seed. A multi-seed run and a Diebold-Mariano test against HAR-RV are pending.
- The covariate ablation (6 / 8 / 14 rows) was run on
chronos-2-small; only the 14-row specification has been evaluated on the 120M checkpoint. - fractional differencing is written although not yet applied on the data