Public research example — internal demonstration

Volatility Term-Structure Strategy

A transparent SPY allocation example combining a lagged VIX term-structure filter, realized-volatility targeting, leverage limits and explicit turnover costs.

This is not a client case study. It is a deliberately simple strategy, built internally, to demonstrate how a transparent quantitative strategy can be documented and reviewed — not evidence of a validated or profitable trading system.

View the repository on GitHub

Strategy overview

The strategy allocates to SPY using two other public instruments as inputs: the VIX index and the VIX3M index.

It holds SPY only when the lagged VIX3M/VIX term-structure ratio is above 1.0 — a simple regime filter for calmer volatility conditions.

When the regime filter is on, exposure is sized to target 10% annualized volatility, using a 20-trading-day trailing realized-volatility estimate, and capped at a maximum of 2x leverage.

A simplified linear cost of 1 basis point per unit of turnover is charged on every change in exposure.

Timing convention

The causal convention is explicit and unit-tested in the repository:

Both signal features — the VIX3M/VIX ratio and realized volatility — are lagged by one trading observation.
Exposure on row t is therefore computed from information available only through t-1.
That exposure is applied to the row-t close-to-close SPY return.
No additional execution shift is used in the published reference implementation.

term_structure_lag[t] = (VIX3M / VIX)[t-1]
realized_vol_lag[t]   = std(SPY returns, 20 days)[t-1] * sqrt(252)
exposure[t]           = 1(term_structure_lag[t] > 1.0)
                        * min(0.10 / realized_vol_lag[t], 2.0)
strategy_return[t]    = exposure[t] * SPY_return[t]
                        - turnover[t] * 1 bp

Snapshot metrics

Sample: 2013-02-01 to 2026-07-17 — 3,384 aligned trading observations.

Strategy vs. SPY buy-and-hold

MetricStrategySPY buy-and-hold

CAGR8.13%14.62%

Annualized volatility10.60%16.86%

Sharpe (0% risk-free rate)0.7910.894

Maximum drawdown-17.17%-33.72%

Total return185.81%525.23%

Additional diagnostics

DiagnosticValue

Annualized active return-6.69%

Information ratio-0.554

Correlation with SPY0.701

Beta to SPY0.441

Average SPY exposure85.1%

Annualized turnover17.15x

Regime on92.6% of observations

Metrics are descriptive history from the committed repository snapshot, not a claim of future performance.

Charts

Line chart of the growth of $1 invested since 2013 in the strategy compared with SPY buy-and-hold, plotted on a logarithmic scale. — Growth of $1

Line chart comparing the strategy drawdown with the SPY buy-and-hold drawdown over time, showing a shallower maximum drawdown for the strategy. — Strategy and SPY drawdowns

Three stacked panels showing the lagged VIX3M/VIX term-structure ratio, the lagged realized-volatility estimate, and the resulting SPY exposure over time. — Lagged term structure, realized volatility and SPY exposure

What this example demonstrates

Explicit timing conventions matter: stating exactly what information each row of exposure uses, and when, is what makes a backtest auditable.
Volatility targeting can reduce realized volatility and historical drawdown relative to a buy-and-hold benchmark.
Lower risk does not automatically imply benchmark outperformance — this example underperformed SPY buy-and-hold on CAGR and total return over the sample.
A clear, like-for-like benchmark prevents risk reduction from being mistaken for alpha.
Reproducible code, machine-readable metrics and derived data make a result inspectable, rather than asking for trust.

Reproducibility

Built and tested with Python 3.12.
Pinned top-level dependencies, listed in the repository’s requirements file.
A standalone analysis script reproduces the metrics, the derived CSV and all three charts from a single command.
Machine-readable metrics are committed alongside the code, together with the exact strategy configuration used.
The derived CSV of aligned returns and exposure is committed and bound to the metrics file by a SHA-256 hash, so the snapshot cannot silently drift from the numbers shown here.
Synthetic-data unit tests cover feature timing, exposure bounds and cost accounting.
A GitHub Actions workflow runs the test suite on every change.

View source on GitHub (opens in a new tab)

Limitations

Yahoo Finance is a convenient public data source, not a point-in-time institutional archive — historical observations may be revised.
VIX and VIX3M are regime indicators; the strategy does not trade VIX products directly.
The backtest does not model bid/ask spreads, market impact, taxes, financing, borrow constraints or intraday execution.
The cost model is a simplified, linear cost per unit of turnover.
The parameters were chosen for demonstration and are not presented as universally optimal.
This is a historical example, not investment advice or a forecast.

Back to case studies