01 FIAM × McGill Asset Management Hackathon 2026

honex.one

Systematic crisis alpha, run by an agentic workflow.

Crisis alphaSystematicAgentic workflowSEC + Carbon Arc data

A market-neutral book that earns from picking stocks, not from where the market goes.

We are a team of five McGill MMF students.

See the live book →Executive summary

Executive summary

Crisis alpha from systematic stock selection: market-neutral, ahead of T-bill + 4% before and after costs, Jan 2021 – Aug 2026.

1We harvest stock-specific mispricing: prices are slow to reflect fundamentals, company filings and real-world demand.
2Five forecasters plus SEC filing and Carbon Arc health signals rank about 3,000 U.S. stocks each month on return beyond market exposure.
3We buy the top ~230 and short the bottom ~230, matched on beta and sector, near-equal weighted, with a daily risk layer.
4Jan 2021 – Aug 2026, net of costs: 11.7% a year vs 7.5% for T-bill + 4%, information ratio 0.56, beta −0.03.

Information ratio

0.71

CAGR

13.0%

Beta to S&P 500

−0.03

Alpha

11.7%

Market sell-offs · monthly backtest, Jan 2021 – Aug 2026

Thesis

Investors underreact to information that is slow to price

Slow to price: shifts in 10-K tone, 8-K filing news, patient claims data

What we look for

Changes in the tone of 10-K filings
8-K filing news and the market's reaction
Medical and pharmacy claims, ahead of revenue

How we trade it

RidgeLightGBMNeural net8-K novelty8-K reaction

Why it is not arbitraged away

1People underreact to information that is slow to price
2Shorting small volatile stocks is expensive
3Stock specific signals remain independent of beta

Data

Data and methodology

What we train on, the alternative data we added, how the forecasters are built, and how the test period was kept clean.

Data

Stock panel: 147 firm characteristics (JKP), CRSP prices, returns and delistings, Compustat fundamentals
Text: SEC EDGAR 8-K filings and SEC filing signals, cut off at 15:30 ET on the last trading day
Carbon Arc: CA0049 Medical & Pharmacy Open Claims (adopted); CA0015, CA0016, CA0035, CA0037, CA0059 tested, not adopted

Carbon Arc CA0049 checks

Coverage: U.S. only, 70–75% of the country; ~200M patients a year
Timing: Weekly; medical claims ~45% complete in week 1, 80% by week 4, so a 4-week lag avoids look-ahead
History: Medical 2015+, pharmacy 2015+ (PBM 2017+): enough to train before 2021
Panel drift: Clearinghouse pharmacies fell from ~38K to ~20K a year: so the signal should use growth rates, not raw counts

Model and splits

Equal-weight ensemble of five forecasters: ridge with a literature prior, LightGBM, a pre-trained neural net and two 8-K signals.

Pre-training	1985–2014	Neural net only
Walk-forward	Train to Y−3, tune Y−2..Y−1, predict Y	Every model
Validation	2017–2020	All rules and risk settings chosen here
Test	Jan 2021 – Aug 2026	Frozen, out of sample

Construction

Portfolio construction

How forecasts become positions: beta and sector neutrality, near-equal sizing, the short-side floor and the mandate constraints.

Data→

Five forecasters→

Ensemble score→

Neutralize→

Match on risk→

Execute

20 risk groups · low beta → high beta

Each square is a stock (sample)

Agent

Agent evidence: agentic workflow for factor-risk monitoring

An AI agent audits the pipeline and monitors exposures. It flags issues for the team and never produces forecasts, positions or trades.

Risk agent scan

Current book + latest Arcana exposures · flags go to the team, the agent never trades

·PositionsMandate 200–500

·Gross exposureMandate ≤ 200%

·Net exposureMandate ±50%

·Largest positionCap 4.9% · avg 0.42%

·Predicted betaTarget ≈ 0 (Arcana)

·Momentum loadingWithin 68-mo range +0.11 to +0.52

·Mid-cap loadingWithin 68-mo range −0.30 to −0.09

·Stock-specific risk shareArcana predicted, 31 Aug 2026

Press Run scan to check the book.

Quant research development

Agentic workflow

1Team idea
2Agent tests it on dev data
3Written rule + code
4Placebo backtest
5Independent audit

What it caught

Exit-date bug8-K survivorship biasLook-ahead leaksCost model

Counterfactual: the agent never sets positions or risk. Every construction and risk decision is a fixed rule, so the live book is the fixed-rule book.

Risk

Risk management: position and book level

Max drawdown, plus the tests that show the returns are not luck.

Losses stayed small and recovered within months.

−4.9%

Monthly max DD

−5.4%

Net of costs

−6.0%

Daily path

1.01%

Daily 99% VaR

HoneX, monthly-4.9%

HoneX, net of costs-5.4%

Arcana max drawdown-4.2%

S&P 500-23.9%

Beta near zero means market sell-offs do not compound into the book's losses.
A drawdown overlay starts cutting risk at a 5% loss from peak.

Live book

Performance: the live book

The current book, marked to market every 5 seconds during the session.

Connecting470 positions as of 2026-08-20 · refreshed every 5 seconds

--:--:--

Portfolio value

$—

Fetching live quotes…

Waiting for intraday data…

Book, % of NAV S&P 500 (SPY)

vs S&P 500 today

—

SPY —

Winners / priced

—

Long leg · 238

—

Short leg · 232

—

Gross exposure

194.5%

$1.656B long · $1.971B short

Net exposure

−16.9%

Beta-matched, not dollar-matched

Top contributorsP&L · move

Top detractorsP&L · move

Quotes from Yahoo Finance, may be delayed. Value = book NAV ($1.864B) + today's P&L on current positions.

Performance

Performance: Exhibit 1

The whole out-of-sample period, Jan 2021 – Aug 2026, net of costs. Three numbers matter most for this mandate.

11.7%

Alpha

1.22

Sharpe ratio

0.56

Information ratio

Metric	Units / basis	Strategy	T-bill + 4%	S&P 500
Ending value of the endowment	US$ billion	1.874	1.507	2.252
Cumulative return	%	87.4%	50.7%	125.2%
Annualised return	% per year	11.7%	7.5%	15.4%
Annualised volatility	% per year	7.5%	0.1%	16.6%
Alpha vs S&P 500	% per year (s.e.)	8.3% (3.4%)	n/a	n/a
Beta vs S&P 500, monthly	coef. (s.e.)	−0.029 (0.070)	n/a	1.00
Beta vs S&P 500, daily	coef. (s.e.)	−0.011 (0.024)	n/a	1.00
Sharpe ratio	annualised	1.10	n/m	0.75
Information ratio vs T-bill + 4%	annualised	0.56	n/a	n/a
Maximum drawdown	%	−6.5%	0.0%	−24.5%
Positions held	avg / min / max	467 / 449 / 499	n/a	n/a
Gross exposure	avg / max, % of capital	187.4% / 199.3%	n/a	n/a
Net exposure	avg / min / max, % of capital	−8.1% / −28.5% / +15.3%	n/a	n/a
One-way turnover	per rebalance / annualised	~18% / ~220%	n/a	n/a

Jan 2021 – Aug 2026, net of costs, daily (alpha and daily beta with Newey–West s.e.). IR gross of costs 0.71. n/m = not meaningful. Turnover estimated.

Arcana

Independent verification: Exhibit 2 (Arcana)

The key points from the Arcana US risk model, 1 Jan 2021 – 31 Aug 2026. Use ← → to flip through.

Exhibit 1 · Idiosyncratic vs factor1 / 11

71% of the return came from factor exposure

Most of the return is factor exposure, mainly style. The residual — what factors cannot explain — added +8.96%.

Arcana value	Full window
Total return	+30.56% ($643.6M)
Factor return	+21.60% ($465.8M)
Residual return	+8.96% ($177.8M)
Residual IR	0.54
Factor Sharpe	1.14
Residual hit rate	42.8%
Residual win / loss	1.49

Total return +30.56%

71%

29%

Factor return

+21.60%

Residual return

+8.96%

1.14

Factor Sharpe

0.54

Residual IR

1.49

Residual win / loss

Residual is what the Arcana US factors cannot explain.

Arcana, New w pharm, Arcana US risk model, historical mode, returns on gross market value. Residual vs. Factor, 1 Jan 2021 – 31 Aug 2026.

Exhibit 2 · Decomposition2 / 11

A short mid-cap tilt earned the most

Style factors explain +23.33%, led by mid cap (+12.13%) and momentum (+6.26%). Components add to the +30.56% total.

Source	Return
Style factors	+23.33%
Mid cap	+12.13%
Momentum	+6.26%
Earnings yield	+4.11%
Size	+3.49%
Volatility	−3.00%
Industry	+1.48%
Beta factors	−3.21%
Residual	+8.96%

Build-up to +30.56%

Style factors

+23.33%

Industry

+1.48%

Beta factors

−3.21%

Residual

+8.96%

Total

+30.56%

Mid cap and size together add +15.6%: the book was short mid caps throughout (loading −0.09 to −0.30).

Arcana, New w pharm, Arcana US risk model, historical. Return Decomp › Full detail, 1 Jan 2021 – 31 Aug 2026.

Exhibit 4 · Concentration3 / 11

Risk is spread across hundreds of names

With 469 positions and 342 effective bets, no small group of names drives the book's risk.

Arcana value	31 Aug 2026
Top 10, share of total risk	14.6%
Top 10, share of stock-specific risk	14.8%
Top 10, share of skill risk	23.1%
Effective independent bets	342
Long / short	175 / 167
Largest position	0.50% of gross

Top 10 names, share of risk

14.6%

of total risk sits in the top 10

Largest position: 0.50% of gross

469 positions■ 342 effective independent bets (175 L / 167 S)

Arcana, New w pharm, Arcana US risk model. Concentration Level (Gross, top 10), book held on 31 Aug 2026.

Summary table4 / 11

Risk snapshot at the end of the window

Forward-looking predictions from Arcana's risk model for the book held on 31 Aug 2026. VaR is a percent of gross.

Arcana value	31 Aug 2026
Predicted volatility	3.56% / yr
Daily VaR (95% / 99%)	0.37% / 0.52%
Annual VaR (95%)	5.86%
Predicted beta	−0.030
Stock-specific share of risk	24.5%
Effective independent bets	342
Positions	469 (237 / 232)
Gross	$3,617M
Net, share of gross	−9.0%

Value at risk on $3,617M gross

0.37%

Daily VaR, 95% ≈ $13M

Where predicted risk comes from

Factor 75.5%Stock-specific 24.5%

3.56%

Predicted vol / yr

−0.030

Predicted beta

342

Effective bets

Arcana, New w pharm, Arcana US risk model, GMV basis. Portfolio views › Summary table, as of 31 Aug 2026.

Exhibit 3 · Long vs short spread5 / 11

Both legs add stock-specific return

The long leg carries 72% of the residual (+6.46% vs +2.50%). Shorts are hedges with some skill, not the main source of alpha.

Arcana value	Full window
Long P&L	+62.52% ($1,346.0M)
Short P&L	−31.96% (−$702.4M)
Total	+30.56%
Long residual	+6.46%
Short residual	+2.50%

Contribution by leg

Long P&L$1,346.0M

+62.52%

Short P&L−$702.4M

−31.96%

Total$643.6M

+30.56%

72%

Residual from the long leg

28%

Residual from the short leg

Shorts lose money in a rising market by design. After removing factors, both legs are positive.

Arcana, New w pharm, Arcana US risk model, historical. Long vs. Short, 1 Jan 2021 – 31 Aug 2026.

Exhibit 5 · Earnings P&L6 / 11

Most return is earned outside earnings season

This is not an event-driven strategy: 94% of P&L comes from outside earnings season.

Arcana value	P&L, $M
Outside earnings season	+612.7
Earnings season (total less outside)	+36.4
Earnings days	+12.4
±5 days around earnings	+24.0
Total (model-agnostic)	+649.1

P&L by calendar, $M

Outside earnings season+612.7M

Earnings season+36.4M

94%

of P&L earned outside earnings season (total $649.1M)

Not an event-driven strategy. Model-agnostic total differs slightly from the risk-model total ($643.6M).

Arcana asset contribution, Earnings Season tab, historical, 1 Jan 2021 – 31 Aug 2026.

Exhibit 6 · Style exposures7 / 11

Long momentum, short mid caps, all the way

Momentum, mid-cap and size loadings never changed sign in 68 months: persistent tilts, not timing.

Loading (z)	31 Aug 26	Range
Momentum	+0.38	+0.11 to +0.52
Mid cap	−0.14	−0.30 to −0.09
Size	−0.08	−0.51 to −0.03
Volatility	+0.14	+0.09 to +0.29
Dividend yield	−0.16	−0.20 to −0.04
Earnings yield	+0.13	−0.01 to +0.27
Value	−0.09	−0.15 to +0.06
Growth	+0.06	−0.10 to +0.17

Loading (z), 31 Aug 2026 vs 68-month range

MomentumPERSISTENT●

Mid capPERSISTENT●

SizePERSISTENT●

VolatilityPERSISTENT●

Dividend yieldPERSISTENT●

Earnings yield

Value

Growth

−0.60+0.6

Bar = range over 68 months, dot = latest. Momentum, mid cap, size, volatility and dividend yield never changed sign: persistent tilts, not timing. All 11 GICS sectors held; largest sector net is Consumer Discretionary, −3.1% of NAV (our daily sector file).

Arcana exposure decomposition, Style mode, month-end loadings, Arcana US risk model, 1 Jan 2021 – 31 Aug 2026.

Exhibit 7 · Drawdown8 / 11

Worst drawdown −4.2%; the S&P fell 24.5%

Our daily recomputation gives −4.26%, consistent with Arcana's −4.23%.

Value	Full window
Max drawdown (Arcana)	−4.23%
Peak	13 Oct 2022
Trough	4 May 2023
S&P 500 max drawdown	−24.5%
S&P trough	12 Oct 2022

Drawdown from peakCompare S&P 500

−4.23%

HoneX, trough 4 May 2023

−24.5%

S&P 500, trough 12 Oct 2022

The book's worst stretch began 13 Oct 2022, the day after the market bottomed, as stocks recovered.

Arcana performance summary (max drawdown) and Arcana daily total returns, compounded; S&P 500 daily total return on the same dates.

Exhibit 8 · Sharpe ratio9 / 11

Sharpe 1.22 overall, 0.54 on stock-specific return

Arcana's Sharpe uses its own risk-free convention; the headline information ratio vs T-bill + 4% is computed separately.

Arcana value	Full window
Sharpe ratio	1.22
Sortino ratio	1.89
Factor Sharpe	1.14
Residual IR	0.54
Residual Sortino	0.89
Realized volatility	3.99%
Factor volatility	2.92%
Residual volatility	2.65%

Annualized, full window

1.22

Sharpe

1.89

Sortino

1.14

Factor Sharpe

0.54

Residual IR

0.89

Residual Sortino

Total Sharpe (1.22) is now above factor Sharpe (1.14): the stock-specific return adds to risk-adjusted performance.

Arcana, New w pharm, Arcana US risk model, historical, annualized. Summary statistics, 1 Jan 2021 – 31 Aug 2026.

Exhibit 9 · Hit rate10 / 11

Half the positions made money, with bigger wins

A win/loss ratio of 1.38 means the book only needs to be right 42.0% of days. Drag the hit rate to see the edge.

Arcana value	Full window
Positions that made money	49% (1,000 / 2,032)
Longs / shorts that made money	58% / 41%
Daily hit rate	49.6%
Win / loss ratio	1.38
Selection P&L	+32.5%
Sizing P&L	−2.0%
Residual: selection / sizing	−16.0% / +25.0%

49%

Positions that made money

58%

Longs

41%

Shorts

Daily hit rate × win/loss (1.38)break-even 42.0%

35%actual 49.6%60%

Daily hit rate 49.5%Edge per unit risked +0.179

Selection vs sizing P&L

Selection

+32.5%

Sizing

−2.0%

1,000 of 2,032 names made money. On stock-specific return, sizing (+25.0%) did the work while picking alone lost (−16.0%).

Arcana, New w pharm, Arcana US risk model, performance summary and Total L/S P&L by name (a name counts on each side it was held), 1 Jan 2021 – 31 Aug 2026.

Exhibit 10 · Turnover11 / 11

About 1.3× gross traded per year

Within the 100–400% range the guideline gives for active quant equity.

Arcana value	Full window
Turnover, annualized	1.27×
Traded per year	$3,379M
Average gross	$2,669M
Holding period (approx.)	~199 trading days

Turnover vs the 100–400% guideline band

100%

400%

0%500%

1.27×

Annualized turnover

$3,379M

Traded per year

~199d

Holding period

What if average gross were $2,669M?

$1,500M$4,000M

Holding period approximated as average gross ÷ annual traded value (0.79 years ≈ 199 trading days).

Arcana performance summary, historical, 1 Jan 2021 – 31 Aug 2026. Portfolio Stats › Turnover.

Use ← → keys, swipe, or the numbers to move between slides.

Discussion

Costs, discussion and what we learned

Costs, how the book behaved, and what we learned.

Trading cost calculator

Pick a stock size and trade size.

Trade size $10M

20bp

Cost per trade

Flat 20bp

$20,000

Cost of this trade

0.50%

Borrow / year

If shorted

Micro and nano caps use the high-low spread estimate, capped at 300bp.

Our costs

20 bp on every trade
High-low spread estimate for the smallest stocks
0.5% a year to borrow our shorts

13.1%

a year before costs

→

11.4%

after costs, above the benchmark

Market neutral

β −0.03

Stayed market neutral, and stayed up through the 2022 rate shock and the April 2025 tariff crash

Filing signals did what we hoped and raised the information ratio

Mid-cap tilt

Mid cap +12.1%of +30.6% total

A good share of our return came from a steady tilt away from mid caps, not just stock picking

Arcana return decomposition

Carbon Arc health data

IR 0.50 → 0.56

Carbon Arc healthcare claims improved returns once added to the model

A template for bringing in other alternative data

More detail ▾

What we learned

Labelling every 8-K as good, bad or neutral news did not predict the market reaction
So we changed how we analyse the filings
Arcana: a good share of return came from a steady tilt away from mid caps, not just stock picking
Carbon Arc health data improved returns and showed how other data sources could be used

With more time

New markets

CanadaTSX
Same time zone and similar accounting data: the easiest first step.
EuropeSTOXX 600 exchanges
Deep universe for a second, lowly correlated book.
JapanTSE Prime
Large cross-section and governance-reform tailwinds.

Deployment for pensions

A market-neutral sleeve with beta ≈ 0 diversifies a pension's equity-heavy book.
Near-equal weights and a 230 + 230 structure keep capacity and liquidity transparent.
Every rule is pre-registered and auditable: a single program rebuilds the book from raw data.

For the portfolio manager

Agents as analysts: quote-verified 8-K events reviewed against each month's strongest picks.
Short-squeeze control: a crowding-aware short screen aimed at the 2024–2025 losses.
Stage-two alternative data plugs into the same leakage canary and placebo tests.

Q&A

Thank you for your time.

Questions?

Thank you to our sponsors