Metrics Reference

quantammsim provides ~30 financial metrics for training objectives and evaluation. All metrics are computed by _calculate_return_value(), which accepts the return_val string from the run fingerprint.

Note

The default training metric is daily_log_sharpe, not sharpe. This uses log returns and daily periodicity, which is more numerically stable for gradient-based optimisation.

Return Metrics

return_val	Description	Differentiable
returns	Total cumulative return: final_value / initial_value - 1	Yes
annualised_returns	Returns annualised by simulation length: (final / initial)^(365*1440 / T) - 1	Yes
returns_over_hodl	Return relative to holding the initial reserves	Yes
annualised_returns_over_hodl	Annualised return relative to holding the initial reserves	Yes
returns_over_uniform_hodl	Return relative to a uniform (equal-value) hold of all assets	Yes
annualised_returns_over_uniform_hodl	Annualised return relative to uniform hold	Yes

Risk-Adjusted Metrics

return_val	Description	Differentiable
daily_log_sharpe	Daily log-return Sharpe ratio, annualised via sqrt(365) (default)	Yes
sharpe	Annualised Sharpe ratio from minute-resolution arithmetic returns	Yes
daily_sharpe	Daily arithmetic-return Sharpe ratio, annualised via sqrt(365)	Yes

Drawdown Metrics

return_val	Description	Differentiable
greatest_draw_down	Maximum peak-to-trough drawdown from initial value (negative)	Approx
weekly_max_drawdown	Worst maximum drawdown across non-overlapping weekly chunks	No (argmax)
calmar	Calmar ratio: annualised return / max drawdown	No
sterling	Sterling ratio: annualised return / average of chunk drawdowns (monthly chunks)	No
ulcer	Negated Ulcer Index: RMS of percentage drawdowns from running peak (monthly chunks)	Approx

Value at Risk Metrics

VaR metrics are available at daily (daily_) and weekly (weekly_) frequencies, at 95% and 99% confidence levels. The _trad suffix uses end-of-period (close-to-close) returns; the non-_trad variant uses all intraday returns within each chunk.

return_val	Description	Differentiable
daily_var_95%	5th percentile of intraday returns (daily chunks)	No (sort)
daily_var_95%_trad	5th percentile of daily close-to-close returns	No (sort)
daily_var_99%	1st percentile of intraday returns (daily chunks)	No (sort)
daily_var_99%_trad	1st percentile of daily close-to-close returns	No (sort)
weekly_var_95%	5th percentile of intraday returns (weekly chunks)	No (sort)
weekly_var_95%_trad	5th percentile of weekly close-to-close returns	No (sort)
weekly_var_99%	1st percentile of intraday returns (weekly chunks)	No (sort)
weekly_var_99%_trad	1st percentile of weekly close-to-close returns	No (sort)

RAROC and ROVAR Metrics

RAROC (Risk-Adjusted Return on Capital) divides annualised total return by annualised VaR. ROVAR (Return Over VaR) annualises per-chunk returns independently, averages, then divides by annualised VaR.

return_val	Description	Differentiable
daily_raroc	RAROC using daily intraday VaR (95%)	No
weekly_raroc	RAROC using weekly intraday VaR (95%)	No
daily_rovar	ROVAR using daily intraday VaR (95%)	No
weekly_rovar	ROVAR using weekly intraday VaR (95%)	No
monthly_rovar	ROVAR using monthly intraday VaR (95%)	No
daily_rovar_trad	ROVAR using daily close-to-close VaR (95%)	Approx
weekly_rovar_trad	ROVAR using weekly close-to-close VaR (95%)	Approx
monthly_rovar_trad	ROVAR using monthly close-to-close VaR (95%)	Approx

Special / Diagnostic Metrics

return_val	Description	Differentiable
value	Full value-over-time array (not a scalar objective)	N/A
reserves_and_values	Dict with final_reserves, final_value, value, prices, reserves	N/A

Choosing a Training Metric

For gradient-based training:

• ``daily_log_sharpe`` (default) – Recommended. Log returns are more numerically stable, and daily periodicity avoids annualisation artifacts.

• ``sharpe`` – Classic choice. Works well but can have gradient issues with very short or very long simulation windows.

• ``calmar`` and ``sterling`` – Drawdown-aware, but contain non-differentiable operations (argmax). Gradients are approximate.

For Optuna-based optimisation (gradient-free):

• Any metric works, since Optuna doesn’t need gradients.

• ``calmar`` and ``weekly_max_drawdown`` are popular for robust strategy selection.