QUARCC Purple: A Market-Neutral US Equity Fund for the McGill-FIAM 2026 Final

QUARCC Purple is a simulated $1 billion US equity fund built for the final round of the McGill-FIAM Asset Management Hackathon 2026. The mandate is market-neutral and the benchmark is the 3-month T-bill plus 4% a year.

From January 2021 to August 2026, after modelled costs, the fund grows to $1.467 billion: a 7.00% CAGR with 3.96% volatility, a Sharpe ratio of 0.91, a maximum drawdown of −2.99%, and an S&P 500 beta of 0.012. It does not beat its benchmark. The information ratio is −0.11 with a standard error of 0.42.

The code is public in the QuarccPurple repository. There is also a fact sheet, a project site, and a paper (PDF) with the full method and references.

Mandate

The fund must hold 200 to 500 names, with gross exposure of at most 200%, net exposure within ±50%, a 5% single-name limit, sector net exposure within ±20%, and no short below a $500 million market capitalisation. We add tighter limits of our own, including a $3 billion floor on new positions.

Forecasting

Encompassing LASSO. The forecaster follows Han, He, Rapach, and Zhou (2019). Each month it fits a value-weighted LASSO on all predictors jointly and a second LASSO that selects among univariate forecasts. A Chong and Hendry (1986) encompassing regression then sets the weight between them:

$$\hat{y}_{i,t} = (1 - \bar{\lambda}_t)\,\hat{y}_{i,t}^{\text{S}} + \bar{\lambda}_t\,\hat{y}_{i,t}^{\text{C}}.$$

The target is the centred cross-sectional rank of each stock’s CAPM beta-adjusted return, so the model is not rewarded for simply loading on high-beta tails.

Anomaly augmentation. To the 147 characteristics of Jensen, Kelly, and Pedersen (2023), we add three signals:

• A mispricing composite after Stambaugh, Yu, and Yuan (2012), averaging signed ranks of ten anomalies available in the panel.
• A walk-forward LightGBM forecast of next-quarter earnings surprises.
• Short interest, used only once its settlement date is public.

The US Long-Short Portfolio takes a different route through the same panel. It finds that forecast skill sits heavily in small-capitalisation shorts, which this mandate largely excludes, and that contrast is worth reading alongside this fund.

Market-state dial. A $3 \times 2$ grid of correlation regime and trailing HML sign scales forecasts by how well the model has done in similar states, shrunk toward its overall record with $N_0 = 12$:

$$\rho_t = \mathrm{clip}\!\left(\frac{n_{\text{cell}}\,\overline{\mathrm{IC}}_{\text{cell},t} + N_0\,\overline{\mathrm{IC}}_{\text{all},t}}{(n_{\text{cell}} + N_0)\,\overline{\mathrm{IC}}_{\text{all},t}},\, 0,\, 1.5\right).$$

At each mid-month formation, nine price signals are refreshed using the latest 21 days of returns, with coefficients held fixed.

Risk

The covariance combines a market factor, 11 GICS sectors, five styles, and five residual principal components. Factor covariance is extended to the holding horizon $H$ with lagged autocovariances:

$$\widetilde{\mathbf{F}}^{(H)} = H\boldsymbol{\Gamma}_0 + \sum_{j=1}^{H-1}(H - j)\big(\boldsymbol{\Gamma}_j + \boldsymbol{\Gamma}_j^\top\big).$$

Specific risk comes from a QLIKE race (Patton, 2011) between HAR and EWMA forecasts, split into downside and upside semivariances. Longs are charged downside variance and shorts upside variance. A two-state Markov stress probability then mixes in a stressed factor matrix.

Construction

A book is formed at each month-end and mid-month, 136 in all, and is traded in equal fifths over five sessions with a 10% daily volume limit. The final program trades expected return against risk, beta, borrow, costs, and low feature coverage:

$$\begin{aligned} \min_{\mathbf{w}} \;\; & -\boldsymbol{\mu}_t^{(H)\top}\mathbf{w} + \frac{\gamma}{2}\Big(\mathbf{w}^\top \mathbf{B}_t \mathbf{F}_t^{(H)}\mathbf{B}_t^\top \mathbf{w} + H\sum_i d_{i,t}^{\pm} w_i^2\Big) \\ & + \lambda_\beta\big(\boldsymbol{\beta}_{\text{hedge}}^\top \mathbf{w}\big)^2 + \frac{H}{252}\mathbf{b}_t^\top \mathbf{w}^- + \sum_i \frac{\lambda_u H d_{i,t}}{c_{i,t}}w_i^2 + \mathrm{TC}(\mathbf{w} - \mathbf{w}_{\text{pre}}) + \text{penalties}, \end{aligned}$$

with $\gamma = 0.25$ and $\lambda_\beta = 10$. To meet the name limit and a 10 bp minimum position, we solve the relaxation, keep at most 440 names, lock their signs, and re-solve with Clarabel through CVXPY. Beta is limited on three separate estimators, and sector net is held within ±1%, far tighter than the mandate requires.

Eight local language-model agents read filings, draft commentary, measure crowding, and challenge the book. None of them sets a weight.

Results

Metric	QUARCC Purple	T-bill + 4%	S&P 500
Ending value of $1.000bn	$1.467bn	$1.509bn	$2.046bn
CAGR	7.00%	7.53%	13.47%
Volatility	3.96%		15.10%
Sharpe ratio	0.91		0.70
Sortino ratio	4.21		1.51
Maximum drawdown	−2.99%		−25.43%
Beta to the S&P 500	0.012		1.00
Information ratio vs. T-bill + 4%	−0.11

Year	Fund	T-bill + 4%
2021	6.98%	4.12%
2022	1.34%	6.19%
2023	15.10%	9.45%
2024	11.16%	9.34%
2025	2.08%	8.37%
2026 (to August)	3.62%	5.21%

The book holds 246 to 357 names, with gross exposure averaging 136.5%, and breaches no mandate limit. Of the $467 million net profit, stock positions contribute $258 million, cash interest $240 million, and costs take $30 million. Interest is nearly half of gross profit, which is why the T-bill-based hurdle is the fair comparison.

Limitations

• Benchmark shortfall. The fund trails its hurdle by 53 bps a year, and costs equal 72% of the cumulative shortfall. The information ratio before costs is about zero. A standard error of 0.42 cannot separate a modest edge from none.
• Selection. Forecasts are walk-forward, but construction settings, including the final gross scale, were chosen with 2021 to 2026 visible, so the sample is not a clean holdout.
• Time variation. The fund trailed the hurdle in 2022, 2025, and 2026 to date. Gross averaged 136.5%, well below the 200% limit.
• Simulated execution. No order was placed. Fills, borrow, and impact at $1 billion may differ. Our paper-fund page marks a single frozen 24 August 2026 formation forward with market data. It is not a live record.

Read more

• Paper (PDF)
• QuarccPurple on GitHub
• Fact sheet (PDF)
• US Long-Short Portfolio