Fox Camann. Business Analytics at the University of Denver (Daniels, 4+1 MS). I do quantitative research on sports betting and prediction markets, and founded SF Strategies.
I price player props across four leagues, grade every forecast against the de-vigged closing price and the result, and treat my own numbers as suspects until a pre-registered test clears them.
+ every forecast graded twice: vs the de-vigged close, and vs what happened
+ every forward test pre-registered: one look, sample bar and kill rule written first
+ the rows the system declined are kept, so the filter itself can be graded
+ a guard that cannot fail on bad input is not a guard
- backtests tuned until they worked| process | state | what it runs |
|---|---|---|
| research-system | running | Daily pipeline pricing NHL, NBA, MLB and WNBA player props from free public data. 1,000+ Python modules, ~1,000 test files, 100+ health checks. Private. |
| prereg | 130 registered | Exploratory study, then a frozen spec: one forward look, fixed sample bar, held-out window, kill rule. Most come back null and stay on file. The framework is open source: prereg (hash-pinned specs, one-look enforcement, anytime-valid confidence sequences, a trial ledger for the multiple-testing count). |
| devig | open source | Removes the bookmaker margin from any 2-way or n-way market: proportional, additive, power, Shin and odds-ratio methods. Pure Python, typed, tested. |
| stateedge | live | SF Strategies. Tracks US sportsbook promotions by state and computes the fee-aware hedge on a prediction-market exchange. Next.js, Supabase, Stripe. |
%%{init: {"theme":"base","themeVariables":{"background":"#0d1117","mainBkg":"#161b22","primaryColor":"#161b22","primaryTextColor":"#e6edf3","primaryBorderColor":"#30363d","lineColor":"#00b4ff","fontFamily":"monospace","edgeLabelBackground":"#0d1117","clusterBkg":"#0d1117"}}}%%
flowchart LR
A["ingest<br/>free public stats + odds"] --> B["models<br/>poisson / neg-binomial<br/>marcel priors"]
B --> C["price<br/>power de-vig vs sharp book"]
C --> D["freeze<br/>taken AND declined rows"]
D --> E["grade<br/>vs close + vs result"]
E --> F{"pre-registered<br/>look passes?"}
F -- "no: stays dark" --> B
F -- "yes" --> G["kelly sizing<br/>fractional, shrunk for<br/>parameter uncertainty"]
G --> H["monitor<br/>anytime-valid CI + kill rule"]
H -. "decay" .-> B
style F stroke:#00ff9c,color:#00ff9c
style G stroke:#00b4ff
style H stroke:#7b2ff7
# config.py
EDGE = {
"pricing": ["poisson", "negative-binomial"], # count models for goals, shots, points
"priors": "marcel, 5/4/3 seasons + regression", # last season is not the player
"devig": "power method", # beat proportional on held-out data
"ruler": "de-vigged closing price",
"inference": {"look": "date-clustered bootstrap", # games on one night are not independent
"decide": "anytime-valid CI"}, # I peek daily; fixed-n p-values lie
"staking": "fractional kelly, shrunk", # the edge itself is an estimate
}cat sizing/kelly.py
def kelly_stake(p_model: float, odds_dec: float, bankroll: float,
fraction: float = 0.25, cap: float = 0.02) -> float:
"""Quarter-Kelly with a hard exposure cap."""
b = odds_dec - 1.0 # net payout per unit staked
edge = p_model * odds_dec - 1.0 # expected value per unit
if edge <= 0:
return 0.0
f_star = edge / b # full kelly fraction
f = min(fraction * f_star, cap)
return round(bankroll * f, 2)
# >>> kelly_stake(p_model=0.52, odds_dec=2.06, bankroll=10_000)
# 167.92A system that grades itself with data it produced can fail in ways its own output cannot show. Every dashboard stays green. Some of what I found on my own book, all reproduced in silent-failure-detectors:
| failure | how it hid |
|---|---|
| CLV computed on a slice the system picked | 25% of prop positions had no closing price, and bets taken were missing one more often than bets passed |
| a filter that could never pass | it demanded more edge than the model may output: 0 of 914 positions could become a bet for 16 days, no error |
| a fair coin nobody can check | the randomised test's early assignments re-derive at 50.6%, a coin flip; the record cannot prove the test was fair |
The detectors run unchanged on federal mortgage data and clinical-trial registrations and find the same shape. The claim is about measurement, not returns.
pandas · numpy · pytest · stripe · kalshi
Bodybuilder. The mandatory poses, each drawn in code, each paired with a line from someone who owned it. Click a pose.
Fourteen small games in one HTML file, no libraries. Trading drills (make a market, fair price, the Kelly coin, poker pot odds, card counting), brain games (four in a row, unbeatable tic-tac-toe, mines, chess vision) and sports (powerlifting meet day, PR chase, shootout, home run derby, free throws). Click to play.
