QUANTT - Complex and ambiguous reports are hard to interpret
- Prices take longer to react
- Our edge is interpretation, not speed
Hirshleifer, Lim & Teoh (2009) · You & Zhang (2009)
McGill-FIAM Asset Management Hackathon 2026 · Finals
Quality minus Junk + Caught in 8-K.
A long/short U.S. equity strategy, beta-neutral by design. We own the highest-quality large companies and short the ones whose 8-K filings read worst inside a language model's brain.
Reading Item 1.01 Entry into a material agreement
looking for junk to short
Caught in 8-K
One filing, one short
Companies must file an 8-K within four business days of a material event. The text is long, and investors underreact to soft, negative news, so the price keeps moving for weeks.
Eos Energy Enterprises · Form 8-K
Item 1.01 Entry into a Material Definitive Agreement.
On May 12, 2026, Eos Energy Enterprises Inc. entered into a binding term sheet with … an affiliate of Cerberus Capital Management, L.P., which provides for … the formation of a joint venture …
The investment by the Company in the JV Company is expected to be financed by a rights offering to holders of Eos Common Shares.
In the event of a liquidation of the JV Company, distributions will be made as follows: (a) first to CCM Frontier until it has received a return of its invested capital …
What happened
Our model scored it in the bottom 2%. We went short. The filing describes a joint venture funded by a rights offering to existing shareholders, with the partner paid back first in a liquidation.
−38%
Eos stock, open rebalance to exit rebalance ($8.97 → $5.55)
+28 bp
added to the whole book
Eos Energy, daily close
8-K released 13 May 2026. Shaded: held short from the 1 June rebalance to the 1 July rebalance.
Executive summary
US$1 billion became US$2.16 billion after every cost
1.31
Sharpe ratio
0.13
beta to the S&P 500
14.6%
CAGR
0.82
information ratio vs T-bill + 4%
−5.9%
maximum drawdown
8.2%
volatility
Out-of-sample backtest, January 2021 to August 2026, US$1bn start, after modelled Interactive Brokers commissions, market impact and borrow costs. All 13 competition rules met on every trading day.
Economic intuition
“The harder a filing is to read, the slower its news reaches the price.”
Our fundamental hypothesis. Three bodies of research explain why it pays, and why our edge is interpretation, not speed.
- Institutional and market barriers keep pessimists from shorting
- Prices reflect the more optimistic investors
- Our edge is capturing a risk premium others will not
Miller (1977)
- Profitable, growing, safe and high-yielding firms earn more
- Diversified across sectors
- Alternative data support the fundamentals
Asness, Frazzini & Pedersen (2019) · Cohen & Frazzini (2008)
Short names with negatively interpreted 8-Ks. Long high-quality firms. Beta-neutral.
The brain scan
We read the model's mind, not its answer
Why not just ask a chatbot whether a filing is good news? Because that squeezes a whole filing into one word. Instead we record what lights up inside the model while it reads.
- Split the 8-KEach item is read on its own. FIAM corpus: 373,139 filings from 2015, each dated at filing.
- The model reads itGemma-2-9B-IT, frozen and offline: no prompt, no tools, no internet. We take its residual stream at layer 20 of 42.
- Pull out featuresA sparse autoencoder (Gemma Scope) turns 3,584 numbers into 131,072 named features; about 81 are active per word.
- Map to returns4,096 features chosen on pre-2019 filings feed a ridge regression on next-month returns, refit each January on past years.
- Score each itemThe lowest predicted returns, ranked within GICS sector, are short candidates.
Features that were activated
# = feature number. Labels from Neuronpedia; every feature shown keeps its sign in at least 5 of 6 yearly refits.
Junk · lower expected return
- Financial distress and corporate restructuring#4391
- Bankruptcy and financial advisory roles#54695
- Financial transactions and capital increases#112725
- Medical studies and clinical trials#37547
- Biotech companies and their clinical projects#116114
- Financial information about pharmaceutical companies#49300
Quality · higher expected return
- Company performance and growth metrics#112779
- Revenue and dividend growth#74985
- Awards and recognitions#66512
- Fast-food and restaurant brands#47606
Explainable shorts. Every feature has a published description, so a portfolio manager can see why a name is on the short list: distress, dilution and drug-trial risk.
A patience trade. Our picks keep earning for three months after each monthly rebalance: on average 0.07% above the hurdle on day 1, 1.2% by day 21 and 3.3% by day 63 (t = 7.0). Dense filings are read slowly, so we can trade slowly.
No memory of prices. The model never sees a ticker, a price, a date after the filing, or a question about the future. It reads one section of text, once.
The pipeline
From five data sources to one portfolio
Two lanes, rebuilt every month: the short leg from 8-K filings, the long leg from fundamentals and demand signals. Both feed one optimiser. Follow the walkthrough, or click any stage.
01
Show every stage as a list
Information known through day t forms the book at the close of day t+1.
What we are buying
Quality
Rank on fundamentals first, then let three alternative signals validate and boost them, then aggregate into our 125 long picks.
Fundamental Scorecard
147 fundamentals combined with shrunk maximum-Sharpe weights (Kozak, Nagel and Santosh, 2020), refit each January on past data and ranked within each sector, so the longs are not a sector bet.
Three alternative signals
- Card spending growth (Carbon Arc)
- Customer momentum: buyers' returns
- Supplier momentum: suppliers' returns
Top 125 picks
The four ranks are added in equal parts (as in Asness, Frazzini and Pedersen, 2019). We buy the best 125 of the 500 largest U.S. companies, weighted by the square root of size, no name above 5%.
Card spending growth
"Are shoppers spending more here than a year ago?"
- Signal: log of the last three months' credit-card transactions minus the same three months a year earlier.
- No look-ahead: used one month after the data month (Agarwal, Qian and Zou, 2021).
- Covers 37% of the 500 candidates but 52% of long gross, mostly consumer companies.
- History starts April 2020; ranked across all 500, a missing value scores zero.
Customer and supplier momentum
"Did this company's buyers and suppliers just have a good month?"
- Signal: last month's value-weighted return of the firm's customer (and supplier) industries, own industry excluded.
- Why it works: investors do not trace links between industries, so the news reaches the price late (Menzly and Ozbas, 2010).
- Links come from the public BEA table, so every firm with an industry code is covered.
Information ratio after costs, January 2021 to August 2026
Same engine, with and without the demand signals. Daily, against T-bill + 4%.
Better where we chose it. At the same 12% volatility, adding card spending and the supply chain lifted the 2019–20 validation information ratio from 1.15 to 1.33. Out of sample it held at about 1.0, with lower beta.
Why we run at 8%. The 12% version realised 10.2% volatility and 14.4% over its worst six months. We cut the target to 8%: the out-of-sample information ratio falls to 0.82, but risk sits well inside the rule and at half the market's.
What we are selling
Junk
Four filters take every company we are allowed to short down to 125 that read badly, trade easily and are cheap to borrow.
Out · small caps
Out · good-reading filings
Out · disputed names
Out · costly trades
weighted by how bad the score is
$2.5bn
median market cap of a short
0.15
days of trading volume, median short position
0.37%
borrow cost, % of capital a year
The alternative-data filter: three groups who know more than the filing
Analysts
Analysts talk to management and revise slowly, and prices are slow to follow their revisions (Chan, Jegadeesh and Lakonishok, 1996). We never short a stock that analysts are upgrading.
Insiders
People sell shares for many reasons but buy for one: they think the stock is cheap (Lakonishok and Lee, 2001). We never short a stock its insiders are buying.
Large funds
When fewer funds hold a stock, the pessimists are leaving and returns tend to be lower afterwards (Chen, Hong and Stein, 2002). We do not short stocks funds are piling into.
Combining the portfolio
Picks, weights, one optimiser, liquidity-aware trading
125 long
125 short
Bigger firm, bigger long. Worse score, bigger short. Leg sizes set so the target beta is zero.
≈ 0Beta to S&P 500
≤ 20%Per GICS sector, net
≤ 8%Expected vol
≤ 200%Gross (net ≤ ±50%)
≤ 5%Any single name
≥ $500mShorts, min size
≤ 1% ADV a day
End-of-month signals and rebalance; execute at the next day's close (T+1). 89% of a month's rebalance is done by day 15.
Beta neutral by design
Leg sizes are set on each stock's 60-month beta, and the optimiser holds the target book's beta at zero every month. Realised daily beta is 0.13, because longs fill in a day and shorts over a week or more.
An 8% volatility target
The rules allow 12–14% and say lower is better. The optimiser caps expected volatility at 8%, about half the market's. Realised volatility was 8.2%.
Checked every day
An independent validator re-checks all 13 rules at every close: 249 to 357 names, gross at most 189%, no name above 5%, every short at least $500m, 10 or more sectors.
The book in numbers
January 2021 to August 2026, measured at every close
| Net exposure, average (range) | +39% (+24% to +50%) |
| Gross exposure, average (max) | 159% (189%) |
| Names held, average (range) | 295 (249–357) |
| Turnover, one-way, a year | 4.5× |
| Average holding period | 144 trading days |
| Costs, % of capital a year | 1.1% trading + 0.4% borrow |
Most-held names
Average weight, % of capital, over all trading days
Long
- JNJJohnson & Johnson2.27
- AAPLApple1.91
- GOOGLAlphabet1.67
- JPMJPMorgan Chase1.59
- CATCaterpillar1.34
Short
- FRHCFreedom Holding−1.01
- CWENClearway Energy−0.69
- ORAOrmat Technologies−0.49
- RILYB. Riley (BRC Group)−0.45
- LAURLaureate Education−0.45
Results
14.6% a year after costs, at half the market's volatility
Every number below is out of sample, after all modelled costs, at a US$1bn fund size where those costs bind.
Growth of $1, January 2021 to August 2026
Monthly, log scale. Our book after trading, impact and borrow costs.
Show as table
0.82
Information ratio, daily
0.83 monthly, against the T-bill + 4% hurdle the rules set.
8.2%
Volatility
S&P 500: 16.6%.
−5.9%
Worst fall from a peak
S&P 500: −24.5%.
Summary statistics
Annualised, daily data unless noted
| Before costs | After costs | T-bill + 4% | S&P 500 | |
|---|---|---|---|---|
| Return a year (compound) | 16.3% | 14.6% | 7.5% | 15.1% |
| Information ratio vs. hurdle | 1.00 | 0.82 | – | – |
| Sharpe ratio | 1.49 | 1.31 | – | 0.74 |
| Sortino ratio | – | 1.98 | – | – |
| Alpha vs. S&P 500, a year (s.e.) | 10.7% (3.4) | 9.2% (3.4) | – | – |
| Beta to S&P 500 (s.e.) | 0.13 (0.01) | 0.13 (0.01) | – | 1.00 |
| Volatility | 8.2% | 8.2% | – | 16.6% |
| Worst fall from a peak | −5.8% | −5.9% | – | −24.5% |
| Months ahead of the hurdle | – | 62% | – | 59% |
Up when the market's down
In 2022 the S&P 500 fell 18%. The book made 15.3%.
+19%
made by the shorts in 2022
−4%
lost by the longs in 2022
+15.3%
the book, after costs
0.13
average beta during the year
Return by calendar year
After costs, against the T-bill + 4% hurdle. 2026 covers January to August.
Show as table
The shorts pay when markets break
Profit by leg, % of capital: the January 2021 short squeeze, the 2022 rate shock, the March 2023 bank failures and the April 2025 tariffs
Show as table
2022 shows the point. Rate rises hit speculative stocks. The longs lost 3.9% of capital, but the junk we shorted re-priced and earned 19.3%.
Steady by design. The book beat the hurdle in 5 of 6 calendar years. It does not try to win a strong bull year like 2023; it earns a steady return with a fraction of the risk.
Costs are fully charged. About 1.1% of capital a year for trading, including market impact, and 0.4% for borrow, at a US$1bn fund size. With the rules' flat borrow fee, the information ratio is 0.63.
Arcana analysis
Checked independently
We uploaded the daily holdings to Arcana (portfolio 34693623, January 2021 to August 2026). Each view is shown exactly as Arcana reports it.
Our own checks agree
Timeline
0
No look-ahead
The validator checked all 419,274 holdings rows: none was traded before the close of the day after its information became public.
5.2%
Not just style tilts
After the Fama-French five factors and momentum, 5.2% a year of alpha remains (t = 2.1), and those factors explain under half of daily variance.
13 / 13
Every rule, every day
All thirteen competition rules pass at every close of the out-of-sample period.
Why our product
Four reasons to invest
Innovation
- A novel way to use an LLM: we read its internal features, not its answers
- Applied to economically and fundamentally proven hypotheses
Portfolio construction & risk
- Monthly signal generation and rebalance
- GICS sector diversification (10+ sectors)
- Liquidity- and market impact-aware
- Targets 8% volatility and zero beta
Alpha generation
- 14.6% CAGR
- 1.31 / 1.98 Sharpe / Sortino
- 8.2% annual realised volatility
- 0.13 realised beta
Continued improvement
- Investigate second- and third-degree links among fundamental and alternative data
- Quantify risk sector by sector to improve robustness
From data to dollars
A journey through our pipeline
You should never invest in something you do not understand. This short film follows a filing from data to dollars.
In one line
Quality minus Junk + is a beta-neutral long/short strategy driven by a language model's brain scan
- 14.6% CAGR
- 1.31 Sharpe
- 1.98 Sortino
- 8.2% annual volatility
- 0.13 beta
- 131k+ features
- 370k+ filings
- 1.6M+ items
- Fundamentals first
- Sector diversified
- Liquidity and market impact aware
- Volatility targeted
The team
Five people, one book
QUANTT is the Queen's University Algorithmic Network & Trading Team. Our FIAM 2026 team spans Queen's, Waterloo and Laurier.