Consumer L/S Book

Ties the four free-data subsector readers (rideshare, airlines, lodging, gaming) into one consumer demand scorecard and a market-neutral long/short tilt.

It is a conviction / sizing input to a discretionary book — not an autopilot.

How it works (plainly) — two separate layers

The alt-data informs the discretionary view; it does not replace it. A short- term demand signal is not a valuation. So the book is thesis-anchored:

1. Demand radar (idea generation). For each subsector, read today's demand temperature from its free signal — how hot/cold demand is running vs its own seasonal trend (a z-score). conviction = temperature × trust, where trust (0–1) is how much that signal has earned: TSA reads Southwest (≈0.9) but barely Delta (0.45, its intl flights are invisible to TSA); NYC reads Uber (0.9) not Lyft (0.2). This is a watchlist, not positions.

2. The book (positions). Direction and conviction come from your fundamental views in config.THESES (−2 strong short … +2 strong long). Two systematic signals only resize each view — never flip it (bounded to ±50%):

   • demand (the radar above) — a confirming reading adds, a contradicting trims;
   • value — a sector-appropriate cheap/expensive read vs each name's own ~5y history: EV/EBIT for the capital-intensive sectors (airlines, casinos, hotels — so leverage and profitability count, which price-to-sales ignores) and price-to-sales for rideshare (where EBIT swings negative). Cheap confirms a long, expensive confirms a short. Free data: price from yfinance, fundamentals from SEC XBRL (EBIT = operating income, which is universally tagged; Q4 derived as annual − 9-month).

   Then beta-neutralize with an SPY hedge.

sized(name)  = thesis × (1 ± up to 50% from demand + value)   # sign locked to thesis
weight(name) = sized / gross ;  + SPY hedge so net beta ≈ 0

So a deeply-undervalued high-conviction long like UBER stays long even when its demand signal is soft — and because the value signal reads it cheap, valuation adds to the long and offsets the soft demand. The earlier version wrongly shorted Uber on a soft demand reading; that was the bug this design fixes.

Example output (live, with UBER set to +2 strong long)

2) THE BOOK  (your THESES, sized by demand + value, beta-neutral)
  name   thesis  demand   value   sized  beta   weight
  UBER     +2.0   -0.24   +0.48   +2.17  1.36  +100.0%   <- soft demand trims, but cheap -> value adds
    + SPY hedge -136% -> net beta +0.00

(Add your other names to config.THESES and they join the book.)

Run

uv run python -m src.book       # 1) demand radar  2) the book  3) macro context
uv run python -m src.backtest   # honest walk-forward test of the systematic signals

Does it actually work? (the honest test)

src/backtest.py builds a point-in-time, dollar-neutral cross-sectional L/S portfolio from each systematic signal every quarter (long what it likes, short what it doesn't), measures the next quarter's return (no look-ahead), and block-bootstraps the mean spread:

demand : +12.3%/yr  90% CI [-0.0%, +23.0%]  hit 70%  n=20  -> not distinguishable from 0
value  :  +1.2%/yr  90% CI [ -8.8%, +10.1%]  hit 62%  n=13  -> not distinguishable from 0

Demand is suggestive (70% hit rate, mostly-positive) but underpowered — its CI still touches zero; value is flat. Neither is a standalone strategy. That is the whole point: the signals are conviction/sizing inputs to a discretionary, thesis-anchored book — not an autopilot. The backtest is here to keep us honest about that, not to advertise an edge.

What it deliberately does NOT do

• No macro overlay. Regime/rates are read judgmentally, not wired into sizing.
• No backtested alpha claim. As the test above shows, the systematic spreads aren't distinguishable from zero on the data we have — so this informs a human's conviction, it doesn't trade itself.

Roadmap

1. ✅ Demand radar across all four subsectors (rideshare, airlines, gaming, hotels).
2. ✅ Beta-neutralized — each name's β vs SPY + an SPY hedge so net beta ≈ 0, with a full risk panel (gross / net-$ / net-β / max).
3. ✅ Lodging wired (hotels MAR/HLT/H) via accommodation employment (provisional).
4. ✅ Thesis-anchored — positions come from your fundamental views; the demand signal only resizes them (±50%), never flips direction. (Fixes the old bug of shorting an undervalued long like UBER on a soft demand quarter.)
5. ✅ Valuation layer (sector-appropriate blend) — EV/EBIT vs own history for the capital-intensive sectors (airlines/casinos/hotels), price-to-sales for rideshare, P/S fallback where EBIT history is too negative. Resizes theses alongside demand — cheap confirms a long, expensive a short — the "other level of research".
6. ✅ Concentration caps — water-fill so no name exceeds 25% and no subsector exceeds 40% of gross, redistributing to names with room. With too few views to diversify, the caps relax (and say so) rather than forcing the book to cash.
7. ✅ Macro context panel — a display-only regime read (yield curve, HY credit spread, unemployment, consumer sentiment, financial conditions from FRED), graded risk-on/off and averaged into a label. By design it never touches sizing — the systematic macro overlay was tested and disproven; macro is read judgmentally only.
8. ✅ Walk-forward signal test — point-in-time, dollar-neutral L/S spread per signal, block-bootstrapped CI (see "Does it actually work?" above). Confirms the systematic spreads aren't distinguishable from zero → conviction inputs, not an autopilot.

Part of a series

The four readers this book consumes: gig · airlines · lodging · gaming.