S&P 500 Backtest — Point-in-Time

loading…

Summary — annualized return (CAGR) by signal

All = tradable equal-weight strategy · Ex top-20% = same strategy with its eventual top-20% winners removed (the "did the rockets carry it?" test) · Mid 50% = interquartile names only. A signal whose Ex top-20% still tops the market is a broad edge, not a lottery.

Quality (0–100) is a composite of return, robustness, annual win-rate, low volatility, and shallow drawdown — everything is ranked by it by default. Click any column header to re-sort (first click high→low, again low→high; Market & SPY stay pinned). Click any signal name for a plain-English explanation.

All strategies vs the market

Relative to the S&P average — is the edge steady or a burst?

Each line = strategy value ÷ S&P-average value, rebased to 100. Rising = beating the market, flat = keeping pace, falling = lagging. A straight climb is a durable edge; a flat line with one jump is a burst. The grey line at 100 is the market.

Overlay ideas — test add-on tactics on any strategy

Click any strategy tile below to load it here (the simulator stays pinned as you scroll the gallery), then pick an overlay idea: 💧 Buy the dip (contribute extra when it dips — fairly compared to investing the same dollars evenly), 📈 Momentum tilt (each month overweight the basket's stocks that rose ≥5% last month), or 📉 Mean-reversion tilt (overweight last month's losers). Each shows the effect and a robustness read — and the honest lesson is that most of these add little; toggle strategies and windows to see which actually hold up.

Each signal — click a tile to load it into the simulator above

Solid = strategy (all) · dashed = ex top-20% · dotted = mid-50% · grey = S&P 500 equal-weight average. Log scale. Year row: green = beat the S&P that year, red = lagged (* = partial). Click a tile to analyze it in the contribution simulator; click the title for its plain-English explainer.

This opens on our strategy — the merged, equally-weighted roster the bot actually buys. A company counts as founder-led when the founder still runs it (CEO / Executive Chairman) or holds control. Below: each stock's growth over the chosen window (rebased to 100 at the window start; recent IPOs begin when they listed).

The orange Our strategy button is the live roster — everything else is comparison. Every line here is founder-led; the buttons choose which universe the average is computed inside, so you can see whether the signal travels beyond the S&P 500 into mid-caps, small-caps, the Dow, or value and dividend names. Each average is that universe's founder-led companies held equal-weight, rebalanced monthly. The indices themselves aren't plotted; S&P 500 index is available as a plain market reference. The roster below follows whichever universe you switch on last. Turn Individual stocks off to read the averages cleanly: one outlier (NVDA is ~+14,800% over 10 years) otherwise flattens everything against the axis. Use the checkboxes in the table to drop names from the average — unticking the biggest winners is the honest test of whether the edge is broad or just a couple of rockets. Each line is rebased at its own first month in the window, so an average whose companies mostly listed recently isn't directly comparable to one running the whole span. The long-window numbers are hindsight, not a result: it applies today's roster backward, so it holds companies picked because they already won, and never holds the founder-led names that failed along the way. Read the 1- and 5-year windows as the honest comparison; the Strategies tab has the point-in-time version that only ever holds what was knowable at the time.

Same plan, one index at a time

$1,000 a month into only the founder-led names of each index, equal-weight, over the window selected above. Same money, same rules, different starting universe. Watch the "Not in S&P 500" column: where it reads subset, that universe's founder-led names are all already in the S&P 500 roster — so a higher number there is concentration, not a different strategy, and it comes with the volatility of holding fewer names. A universe only offers genuinely new companies when that count is above zero. The two baseline rows at the bottom are the same $500/month with no founder screen at all — equal-weight is the like-for-like comparison, since every founder-led average above is equal-weighted too.

The roster —

Click a column header to sort · click a row to highlight that stock on the chart above (click again to reset). Return is over the selected window (or since IPO if the company listed mid-window).

The founder-led bot — $500 twice a month (the 1st and the 15th, or the next trading day), split equally across one merged roster of the founder-led names in the S&P 500 and the Nasdaq-100, on an Alpaca paper account. Every run re-audits the rosters first, sells anything that stopped being founder-led, and recycles the proceeds into the current names. The chart starts on the date of the first real trade; the simulated line is the same plan from that same month, for comparison.

Holdings — —

Click a column header to sort. Cost basis and P/L come straight from the account.

The weekly put-spread strategy from the hackathon research: every week, sell a SPY put spread 3% below the market (short strike at 97% of spot, long strike one width lower — about $5 on today's SPY), collect the credit, and hold to the next week's expiry. It wins the week whenever SPY doesn't fall more than ~3%; when it does, it loses up to the full width. Start with $100,000, add $500 on the 1st and the 15th, and compare with the same money in the S&P 500 (dividends reinvested).

Risk per week is the share of the account that would be lost if SPY finished below the long strike — the only real lever in this strategy, because nearly every week is a small win and the losses are full-size. Payroll weeks (the jobs report lands on the first Friday of the month) looked dangerous in the 2024–26 real data — but over the full 30 years they were not worse than other weeks, and sitting them out just forgoes a quarter of the premium. The toggle is here so you can see that for yourself. The chart is log-scale; the S&P line is the same $100,000 and the same deposits, bought on the same days. Real option prices from Feb 2024; earlier weeks are modeled (see the note under the table).

Every window at a glance

All five windows at the risk and payroll rule selected above, ending on the latest completed week. Return/yr is money-weighted (it accounts for the deposits arriving over time), so it is comparable across the two columns. Max DD is the deepest peak-to-trough fall in account value.

Year by year

Calendar-year return of the strategy at the selected risk (no deposits, so it's a pure return) against the S&P 500. Green = beat the index that year, red = lagged. Look at 2008, 2020 and 2022 before deciding this is easy money.

The settings picked for the Alpaca AI Trading Agents Hackathon (Aug 28 – Sep 4, 2026): SPY put credit spreads 3% below the market, $5 wide, one-week expiry bought back for pennies on the last morning, 70% of the account at risk, no stop. Below: a single $100,000 with no deposits, run through those exact settings over the last 1, 3, 6, 12, 24 and 36 months, against SPY.

Nothing here changes the odds — roughly 93–95% that a week ends green, 5–8% that SPY drops more than 3% and the week loses the whole amount at risk. Size only changes what each outcome is worth: at 70% a normal week is about +5% and a bad one is −70%. The red dots are those bad weeks. Read the 24- and 36-month panels before the 12-month one — the last twelve months had no losing week at all, which is the regime, not the rule.

The same six windows in numbers

Contest result

Contest paper trades — live

The agent's real activity on its own Alpaca paper account (fresh $100,000, options level 3). It writes this panel after every gate, entry, half-hourly mark and buy-back, and redeploys the site. The gate is Claude reading the chain through the Alpaca MCP server and the news before each entry — its reasoning is in the journal below.

Agent journal

The buy-the-dip rule from the 2015–16 TradeStation research, rebuilt on the E-mini S&P 500 (ES): when ES trades X% below the session's open, buy once at that level. No stop. Sell at the close of the Nth session — or at a +5% target if you switch it on. One position at a time. Start with $100,000; the whole account earns T-bill interest every day (futures are margined, so the cash never leaves), and a position is leverage × equity of ES exposure. Compared with the same $100,000 in ES buy-and-hold on the same T-bill accounting.

The defaults are the robust version: a 2% dip, taken only when the index closed above its 200-day average the day before, held 30 sessions, no target. It is positive in every decade since 1993, and through the 2000–02 and 2008 bear markets its drawdown was about a third of the index's — but it does not beat buy-and-hold: it is in the market only a third of the time, and in 2020 it took the same hit as the index because there is no stop. Read it as a lower-exposure rule with a real per-trade edge, not a replacement for holding the index. The original 4% dip with no filter is one click away — it fires only in crashes (75 times in 33 years, half of them in 2008 and 2020), so its results are a handful of trades. Every stop-loss tested, on daily bars 2000–26 and on minute bars 2012–26, made the rule worse: on a day the market is already down 2–4%, a 1–2% stop is hit routinely before the bounce. Leverage multiplies the return and the drawdown — at 3× a 2020 or 2008 entry loses most of the account.

Every window at a glance

All five windows at the settings selected above, ending on the latest session. /yr is the compound annual growth rate. Max DD is the deepest peak-to-trough fall in account value, marked to market daily. In market is the share of sessions with a position on; the rest of the time the account just earns T-bills.

Year by year

Calendar-year change in account value at the selected settings against ES buy-and-hold. Green = beat the index that year, red = lagged. A year the rule sat out entirely shows the T-bill return.