Regime-Adaptive Student-t GARCH Volatility Targeting for the NEPSE Index

Strategy Card

Regime-Adaptive Student-t GARCH Volatility Targeting for the NEPSE Index — strategy card

Spec ID: spec-regime-adaptive-garch-volatility-targeting-us-etfs-1786660810 · Generated: 2026-08-13 23:25 UTC

Cluster: Volatility Targeting · Sub Cluster: Regime-Adaptive Student-T Garch Allocation

One-line description

The strategy holds a long NEPSE Index exposure scaled to a 12% annualized volatility target using one-day Student-t GARCH or EGARCH forecasts. It shortens the estimation window and halves exposure when point-in-time stability diagnostics flag an unstable regime.

Why this trade exists

Equity-index volatility is clustered, persistent, fat-tailed, and potentially asymmetric. A constant notional position therefore takes more risk during turbulent periods and less risk during calm periods. Scaling exposure inversely to forecast volatility seeks a more stable risk contribution and may reduce drawdowns caused by sudden volatility expansion.

GARCH and EGARCH can perform differently as shock asymmetry and persistence evolve. Rolling forecast-loss selection avoids permanently committing to either specification, while structural-break, forecast-error, and parameter-shift diagnostics address model degradation by using a shorter estimation window and a 0.5 exposure multiplier. The expected benefit is risk stabilization rather than directional return prediction.

Algorithm

[code omitted from public view]

Parameters

Param	Value	Notes
Instrument	NEPSE Index	Single-index, long-only universe
Data / frequency	Daily close	Log returns; daily close rebalancing
Backtest window	2022-01-01 to 2026-08-12	Up to 919 earlier valid returns may be loaded solely for warmup
Forecast horizon	1 trading day	Conditional variance for the next close-to-close interval
Execution	Same-day close, MOC	Signal uses information through close t
Conditional mean	AR(3)	Common to every volatility model
Innovations	Standardized Student-t	Maximum-likelihood estimation with hybrid solver
Selection models	GARCH(1,1), EGARCH(1,1)	GJR-GARCH(1,1) is robustness-only
Model-selection loss	60-forecast RMSE	Squared log return versus forecast variance; minimum 20 forecasts for dynamic selection
Stable / unstable windows	919 / 504 returns	Window shortens after an instability trigger
Regime triggers	Recent break, error z-score, parameter shift	Bai-Perron: 5% level and 126-return recency; error: 2.5 over 60 forecasts; shift score: 3.0 over 20 refits
Minimum unstable duration	20 trading days	Release requires all three diagnostics to be untriggered
Volatility target	12% annualized	Uses 252 trading days per year
Regime multipliers	1.0 stable; 0.5 unstable	Applied after inverse-volatility sizing
Position / leverage limits	100% long-only position; 4.0 gross maximum	The single-position cap normally binds before the platform leverage limit
Strategy turnover cost	2 bps one way	Cost equals 0.0002 times absolute target-weight change; gross and net reporting requested
Platform cost schedule	0 slippage; $0.004/share; $1 minimum; 1% maximum commission	Separate backtest-engine settings supplied with the run

Look-ahead audit

#	Concern	Status
1	Close-derived signal executed at the same day's open	Pass — execution is market-on-close, not at the open
2	One-step forecast uses the realized t+1 return	Pass — the forecast is formed through close t and applied to the close-t to close-(t+1) interval
3	Rolling model loss includes outcomes unavailable at selection time	Pass — only completed forecasts and subsequently observed returns enter trailing RMSE and MAE
4	Structural breaks or parameter shifts use full-sample information	Pass by specification — all regime diagnostics are re-estimated point in time through t
5	Pre-backtest history leaks evaluation-period outcomes	Pass — earlier data are restricted to model warmup and do not generate reported-period trades
6	Missing execution prices are synthesized	Pass — affected trades are skipped without forward-filling or fabrication

Caveats / known limitations

• The executable candidate definitions omit ARCH(1), ARCH(2), and ARCH(3), despite the implementation note claiming that they are restored. The paper's Table 4 comparison of those models with GARCH(1,1) therefore does not describe this implementation.
• ARCH-LM tests at lags 5, 10, and 15 are diagnostic only and do not condition estimation or trading, unlike the paper's stated justification for applying ARCH-family models. Any interpretation that this strategy trades only when the paper's ARCH-effect condition holds is invalid.
• Model selection requires a "diagnostically valid" GARCH or EGARCH fit, but the specification supplies no p-value thresholds or failure rules for that status. Candidate eligibility and therefore the traded model path are underdetermined unless the implementation defines those rules externally.
• Two declared but undescribed audit deviations prevent direct reproduction of the paper's fixed 230-forecast experiment, model-specific RMSE and MAE, Mincer-Zarnowitz R-squared values, and full-sample or subperiod persistence and half-life comparisons.
• Dynamic GARCH-versus-EGARCH selection, the 504-return unstable window, regime de-risking, and squared-return operational loss differ from the paper's separate fixed-holdout evaluations. The paper's GARCH RMSE 0.009552, EGARCH RMSE 0.009670, GARCH MAE 0.007521, EGARCH MAE 0.007406, and associated Mincer-Zarnowitz rankings cannot be assigned to the composite forecast traded here.
• The 12% volatility target, long-only sizing, position and leverage caps, cash allocation, and transaction costs are trading-overlay adaptations. They invalidate attribution of this run's return, drawdown, turnover, cost-adjusted performance, or target accuracy to the source paper.
• The raw run output does not report return-distribution moments, GARCH-family coefficients, persistence, half-lives, forecast errors, Mincer-Zarnowitz regressions, asymmetry tests, or detected break dates. It therefore cannot establish whether the paper's predicted NEPSE distribution, parameter estimates, leverage evidence, forecast ranking, or 13 July 2022 structural break are reproduced.
• The run reports 1,150 trades but zero symbols and an empty symbol list. This internal inconsistency, together with the fact that an index may require a proxy or derivative for execution, prevents independent confirmation that the reported P&L is attributable to tradable NEPSE exposure.

Results

The reported backtest produced a 37.30% total return with 9.79% annualized volatility, a 0.83 Sharpe ratio, a 1.11 Sortino ratio, and a -11.44% maximum drawdown. The reported win rate was 73.89% and profit factor 2.59 across 1,150 trades. These headline results suggest comparatively controlled realized risk, but they should be treated as provisional because the output identifies no traded symbols and does not provide the forecast, regime, turnover, or volatility-target diagnostics needed to attribute performance to the adaptive GARCH mechanism.

Run diagnostics

Recomputed here from the run's stored daily returns, because every annualized figure in the table above is derived from the LENGTH of that array (years = n / 252). Only total return and max drawdown do not depend on the row count, so when the recomputed and stored values differ by a common factor, those two are the numbers to trust.

Check	From the run's own returns	Note
Daily-return rows	1,156	0.99x the 1,162 trading days in 2022-01-01..2026-08-12
Distinct dates	1,156	one row per date
Date span	2022-01-03 .. 2026-08-12
Sum of daily returns	37.30%	matches the reported total return
Sharpe from these rows	0.83	stored 0.83
Volatility from these rows	9.79%	stored 9.79%
Max drawdown from these rows	-11.44%	stored -11.44%
CAGR from these rows	7.15%	stored 7.15%

Backtest metrics snapshot

Metric	Value
Total Return	37.30%
Sharpe	0.83
Sortino	1.11
Calmar	0.63
Max Drawdown	-11.44%
Volatility	9.79%
Beta vs SPY	0.45
Win Rate	73.89%
Profit Factor	2.59
Total Trades	1150
Symbols	0