Research

Research before execution.

A run-level record of targets, features, 11-fold expanding-window OOS methodology, positive results, negative ablations and deployment status.

Model registry

Run-level evidence, including negative findings.

Every quantitative result is paired with its run ID and repository source. Research status and production status are separate fields.

Historical research

Model	Target / horizon	Features	Validation	Status	Run ID
XGBoost classifier Gen-1	Next-H1 direction 1 hour	71 point-in-time tabular features	11 expanding-window annual OOS folds, 2016–2026; 7,590,789 OOS bars	INACTIVE	xgboost_20260903_200844
Catalogue metrics Gross Sharpe 1.444 Hit rate 52.1% Methodology Point-in-time tabular features evaluated across annual expanding-window out-of-sample folds. Repository model catalog and run artifacts · private research repository · main
LightGBM classifier Gen-1	Next-H1 direction 1 hour	81 point-in-time features	11 expanding-window annual folds	INACTIVE	lightgbm_20260908_170458
Hybrid XGBoost regression Gen-2	Next-H1 forward log-return 1 hour	103 features: 71 tabular + 32 causal TCN latent features	11-fold walk-forward evaluation	RESEARCH FINDING	xgboost_reg_20260912_160548
Hybrid XGBoost classification Gen-2	Next-H1 direction 1 hour	103 features: 71 tabular + 32 causal TCN latent features	11-fold OOS evaluation, 2016–2026	STAGING	xgboost_20260916_103341
DL-only regression ablation Research	Next-H1 forward log-return 1 hour	Causal TCN representation only	Repository comparison run	RESEARCH FINDING	comparison_20260912_053630
DL-only classification ablation Research	Next-H1 direction 1 hour	Causal TCN representation only	Repository comparison run	RESEARCH FINDING	xgboost_20260917_082540

Threshold research

Net Sharpe by classification threshold

Historical backtest

τ 0.50

-2.86

τ 0.58

+0.69

τ 0.60

+1.85

τ 0.62

+2.96

τ 0.65

+3.99

Run source: Repository model catalog and run artifacts. Historical threshold sensitivity, not a promised operating point.

Target registry

Next-H1 directionImplemented

Next-H1 forward log-returnImplemented

Log next-H1 realized varianceExperiment branch

Triple-barrier targetImplemented for future Phase 2

Volatility family

XGBoostVolatilityModel (primary) · 2 baselines

Target

log_realized_variance (next H1)

Baselines

Persistence baseline · HAR-RV baseline

Completed-bar windows

1 / 24 / 120 completed H1 bars

Limitations

All reported metrics are historical repository-catalogued research/backtest results. They are not independently audited, do not represent live trading, and do not guarantee future performance. Costs, slippage, liquidity, regime change and implementation differences can materially alter outcomes.

Model architecture

Three model tasks feed one downstream decision layer.

Direction and expected return describe the opportunity. The volatility family describes its expected risk, so sizing and risk limits can scale with forecast variance instead of a fixed assumption.

Architecture · not live

Classification

Which way will the next H1 bar move?

Output

Direction probability

Models

Gen-1 XGBoost classifier
Gen-2 hybrid XGBoost classifier

Regression

How large is the expected next-H1 log return?

Output

Continuous expected return

Models

Gen-2 hybrid XGBoost regression

Volatility forecasting

How much will price vary over the next H1 bar?

Output

log_realized_variance

Models

XGBoostVolatilityModel
Persistence baseline
HAR-RV baseline

↓ Decision policy · confidence filter · risk engine · execution (operator-owned)

Volatility model family · experiment/xgb-volatility-h1

One primary forecaster, tested against leakage-safe baselines.

Every variant targets log_realized_variance for the next H1 bar. The baselines exist to show whether XGBoostVolatilityModel adds anything beyond simple persistence and HAR-RV structure.

Research · no results published

3 of 5 reported variants verified. Five volatility model types were reported for this branch. Only three are verifiable from the supplied repository facts; the remaining two are not shown because their names and definitions could not be confirmed.

PRIMARY FORECASTER

XGBoostVolatilityModel

Gradient-boosted tree regression on the log of next-bar realized variance.

Target: log_realized_variance
Horizon: Next H1 bar
Inputs: Point-in-time feature set from the pipeline
Causal construction: Target is the realized variance of the next completed H1 bar; features use only information available at bar close.
Validation status: Experiment branch · no performance results published
Artifact status: Trained artifact not included in repository

models/volatility.py · private research repository

BASELINEDescriptive name

Persistence baseline

Leakage-safe persistence: forecasts next-H1 variance from the most recent completed realized variance.

Target: log_realized_variance
Horizon: Next H1 bar
Inputs: Most recent completed H1 realized variance
Causal construction: Uses completed bars only; no current or future bar enters the forecast.
Validation status: Benchmark implementation · no performance results published
Artifact status: Trained artifact not included in repository

models/volatility_baselines.py · private research repository

BASELINEDescriptive name

HAR-RV baseline

Heterogeneous autoregressive realized-variance model combining short, medium and long causal windows.

Target: log_realized_variance
Horizon: Next H1 bar
Inputs: Realized variance over 1, 24 and 120 completed H1 bars
Causal construction: Causal completed-H1 windows of 1 / 24 / 120 bars; windows close before the forecast bar opens.
Validation status: Benchmark implementation · no performance results published
Artifact status: Trained artifact not included in repository

models/volatility_baselines.py · private research repository

Catalog references (private repository): models/MODEL_CATALOG.md · models/catalog.json. Private research repository — no public links.

Research engine

11-fold expanding-window OOS validation (2016–2026)

For each validation year Y+1, train on all available history from 2002 through Y, then evaluate only on Y+1.

Historical research design

Annual full-retrain sequence

TrainingOut-of-sample

20022008201420202026

F012016

F022017

F032018

F042019

F052020

F062021

F072022

F082023

F092024

F102025

F112026

Fold rule

For each validation year Y+1, train on all available history from 2002 through Y, then evaluate only on Y+1.

Temporal control

Point-in-time features and strict temporal causality prevent future bars from entering training or inference inputs.

LEAN distinction

This research walk-forward design is separate from QuantConnect LEAN backtesting. LEAN status remains NOT YET TESTED.

Expanding window

Train on 2002 through Y, then validate only on Y+1.

Strict causality

Point-in-time features use only information available at the bar timestamp.

Full retrain

The complete model is retrained independently for every annual fold.

Not LEAN

Research OOS validation is separate from QuantConnect LEAN backtesting.

Representation research: a causal dilated TCN autoencoder was explored for 32-dim market-regime embeddings during gen-2 research. The production triad does not use TCN latents — direction, magnitude and volatility are served from gradient-boosted models on tabular features.

Threshold research

Higher thresholds reduced trade coverage and changed net backtest results.

These are historical threshold-study observations, not an operating promise or guarantee.

Historical backtest

Threshold τ

0.50

Net Sharpe

-2.86