📖 Spectrum · Strategy

What it trades, the exact rules, and the evidence behind each choice.

In one breath

cross-sectional momentum · Jegadeesh–Titman (1993) · Carhart 12-1 convention
Every Friday we line up ~80 big, easy-to-trade companies and measure how far each climbed over the past 12 months — ignoring the very last month, because last month's fireworks usually fizzle. The 3 highest climbers get the money, split equally. Anyone with a report card (earnings) due in the next 10 days sits out. Anyone who's actually DOWN over the year sits out too (their slot stays in cash). Next Friday, we re-run the race.

Formally: cross-sectional 12-1 momentum, long-only, concentrated. Winners keep winning for months at a time — one of the oldest and most replicated effects in finance (documented in US stocks since 1993, and in data going back a century). Spectrum buys the top 3 names by close[t−21] / close[t−252] − 1 from a point-in-time universe of the 80 most liquid large caps, rebalanced weekly, net of 5 bps turnover costs.

top 3 nameslong only weekly re-rank (Fri)12-1 momentum earnings-avoid 10dabsolute-momentum filter 5 bps costs modeledpaper only

The Friday checklist — exact rules

each rule earned its place with data, not opinion
#RuleWhy it exists (the evidence)
1Universe = the 80 most liquid of ~133 large caps, measured as of that Friday (delisted names included in history). Point-in-time liquidity screen kills survivorship bias: dead names (SIVB, FRC, TWTR…) are in the backtest until the day they actually died. Today's list can't leak into yesterday's decision.
2Rank everyone by 12-month momentum, skipping the most recent month (12-1). The last month mean-reverts (short-term reversal). A pre-registered sweep of 3/6/9-month windows (trials #15–17) lost to 12-1 on every metric — quarterly momentum collapsed to a −70% drawdown. The graded record lives in the ledger on the Architecture tab.
3Drop any name reporting earnings within 10 days (proxy: SEC quarterly filing dates); next-ranked name fills the slot. Loss autopsy of the 10-year book: the worst weeks were single-name earnings blowups. On a 3-name book one −30% report is −10% of the account. (On a diversified book this rule subtracts value — it's concentration-specific.)
4Absolute-momentum filter: a name whose own 12-1 momentum is negative doesn't get bought; its slot stays in cash. Antonacci-style bear safety. In a broad crash "the best of the losers" is still a loser — cash is the better slot. Rarely fires in bull markets; exists for 2018/2022-style regimes.
5Buy the top 3 equal-dollar; hold one week; repeat. Concentration is where the return comes from. Equal weight avoids fitting weights to the backtest.
6Costs are charged on every change: 5 bps × turnover, symmetric. 13 intraday strategies tested by this project died at realistic costs. Any rule that only works at zero cost is not a strategy — it's an execution assumption.

The report card — 10-year backtest, net of costs

weekly book vs SPY vs QQQ · trade-level stats from the holding episodes
MetricStrategy (top-3, 12-1)SPYQQQ
Loading…
Trade-level stats — one row per holding episode (… episodes; open positions marked to the latest close):
Win %Profit factorPayoff ratio Avg winAvg lossBest trade Worst tradeAvg held
Loading…
Profit factor = gross wins ÷ gross losses (>1 means the wins pay for the losses). Payoff ratio = average win ÷ average loss. With a payoff ratio near 2, even a ~50% win rate compounds strongly — that's the shape of a momentum book: many small losses, a few huge winners.
The honest risks: 3 names ≈ 100% semiconductors lately — the book swings ~1.5–2× QQQ. The backtest window contains the AI/semis supercycle; expect leaner years ahead. Max drawdown was −41% and the worst single week −21% — you WILL live through stretches that feel broken. A −7% year (2022) is the good outcome in a bear market. Weekly discipline matters more than any single pick. Paper trading only.