Research

Before the committee traded a dollar of paper money, its strategy was run against 30 years of market history, from March 1996 to May 2026. This page shows the result, the reasons to distrust it, and the ideas that failed along the way.

Thirty years against SPY

monthly series, both columns measured alike

The locked strategy against buy-and-hold SPY, March 1996 to May 2026
measure	strategy	SPY
annual return	20.0%	10.4%
volatility	24.5%	15.3%
sharpe ratio	0.82	0.68
worst fall	−33.7%	−50.8%
beta to SPY	0.57	1.00
share of SPY's losses it suffered	37%	100%

The strategy returned roughly twice as much as SPY each year, and its worst fall was shallower. The protection comes from holding cash during bear markets. It has costs: the strategy is more volatile than the index, and outside bear markets it holds ordinary stocks.

Return by decade

most of the return is from the weakest data

Annual return by decade
decade	a year	data quality
2000 to 2009	30.6%	includes 2000–2006, where price coverage is 73–80%
2010 to 2019	8.4%	price coverage above 90% from 2011

The 20% headline leans on the 2000s. That is the decade where the history has the most gaps, so the best-performing period is also the least reliable one. In the 2010s, where coverage is good, the strategy made 8.4% a year, so 20% is better treated as an upper bound.

What was tried

four experiments and their results

What should the strategy hold in a bear market?kept
May 2026worst fall −60.7% → −35.2%
The first 30-year run fell 60.7%, bottoming in November 2008. The market signal had fired correctly in December 2007. The damage came from what the strategy bought next: its value screen ranked Genworth, Hartford, E*Trade, Sallie Mae and Pulte as the cheapest stocks available, and they were cheap because they were failing. Bear months now go to cash. The cash rule was chosen from that diagnosis, and no fraction between 0% and 100% was tried, since picking one would have meant fitting it to the same history.
Wait a second month before selling out?rejected
May 2026worst fall −35.2% → −54.8%
Twelve of the thirty bear signals over 30 years lasted exactly one month, and all twelve reversed. Waiting for confirmation would have skipped them and lifted 2010s returns from 8.4% to 10.7% a year. It also gutted the 2008 protection: dead-cat bounces kept resetting the counter, so the strategy stayed invested through most of the crash. The limit is a 35% worst fall; this missed it by 20 points.
Weight positions by inverse volatility?removed
2022–2025−1.54 points a year, Sharpe −0.12
Tested against plain equal weighting over 48 monthly rebalances. The gap widened every year, from $475 in mid-2022 to $12,665 by the end of 2025, with identical turnover. The optimiser was shrinking the volatile smaller companies that produce most of the return.
Rebalance every two weeks?held
July 2026Sharpe 0.93 → 1.16
Biweekly beat monthly over 2021 to 2026, and kept a 0.22 Sharpe lead even with trading costs tripled. Walk-forward testing agreed, at 1.01 against 0.90. The live book still rebalances monthly, because the strategy is locked for the length of its public evaluation and changes wait until it ends.

Worst-fall figures in this log come from the daily series. That is why the locked strategy reads −35.2% here and −33.7% in the monthly table above.

Testing rules

point in time: A quarter's results are only used 90 days after the quarter closes, so the backtest never trades on a number the public had not yet seen.
survivorship: Each date's universe is the S&P 500 as it stood at the latest quarter-end, built from 767 historical members. Companies that later collapsed or were acquired stay in for the dates they belonged, though many have gaps in their price history.
walk-forward: Candidates are scored on test windows they were never tuned on. A strategy that only looks good on the full period does not get locked.
costs: Every trade pays commission and slippage. Cost-sensitivity runs repeat a result at two and three times those costs.
a gate set in advance: The 30-year run has a fixed limit: a worst fall beyond 35% fails, however good the return.

Known limits

not yet fixed

thin early data: Price coverage is 71–73% of the universe in 1996–1999 and only reaches 90% after 2010. Fundamentals are sparse before about 2006.
restated figures: The data provider reports the latest restated figures, so after a restatement the backtest sees numbers that differ from what was filed at the time.
flat cash rate: Cash earns 3% a year throughout these results. That flatters 2008–2015, when rates were near zero, and shortchanges 2022–2024. The engine can now use the historical 3-month Treasury bill rate, and the next re-run will replace these figures.
signal lag: The bear signal waits for SPY to cross its 200-day average. In 2008 it fired two months after the peak, and in 2009 it returned two months after the bottom.
perfect fills: Trades fill at the next open in full. Thin names during a crisis such as October 2008 would not have.

The data behind all of this, and why it cannot be downloaded again, is on the data page.