Daily Bilbo Boxes

Sometimes a stock goes quiet: its daily swings shrink and price coils into a tight range for days or weeks. This study asks a simple question — when price finally bursts out of the top of that quiet range, is it worth buying?

Yes, measurably. Buying the upside breakout of these “compression boxes” made more money than buying the same stocks on randomly chosen days with identical risk controls — our test for “is this skill or luck?” It kept working on 30 stocks that were never used to design the rules. The reverse trades don’t work: when price breaks out the bottom, neither betting on more downside nor bargain-hunting beats a coin flip.

What it feels like to trade: about two out of three trades lose small — typically 1–3%, because the exit plan cuts losers fast — and a handful of trades each year run for months and pay for everything (examples below include +156% in TSLA and +110% in INTC). If small frequent losses bother you, this system will too.

The rules, in one breath: wait for the quiet spell; place a buy order just above its range top, only when the stock is already above its 21-day average; if the trade fires, the worst-case exit is the range bottom; as the trade works, follow it up with the 50-day average and sell when price falls through it. A second, faster version watches during the day to catch ranges that break out the moment they end. And the account is always flat when a company reports earnings — one overnight earnings surprise taught us that lesson below.

What this is not: a prediction. It’s a measurement of what these rules did in the past, before fees, in a simulation. The forward test on a paper account is running now.

Saty Mahajan’s Phase Oscillator measures where price sits relative to its 21 EMA, scaled by ATR. Its compression flag fires when Bollinger bands squeeze inside the ATR envelope — daily ranges shrinking, buyers and sellers agreeing on price. Squeezes end. The question a trader cares about is which way, and whether the resolution is tradeable after honest costs and honest baselines.

The Bilbo box is the high and low of the first 5 compression days. That range converts a fuzzy “squeeze” into two hard prices: break the top → long trigger; touch the bottom first → stand aside (we measured the down-break every way we could think of — it’s dead, see the NO-GO list below). The entry is a resting buy-stop at the box high, gated by the prior close being above the 21 EMA so we only take breaks in names already trending up. Nothing about the entry needs hindsight: the order can physically rest at the broker the moment the box locks.

Initial stop at the box low — if the breakout was false, the box itself says so. From there the stop only moves up, ratcheting to yesterday’s 50 EMA each day. There is no profit target: 65% of trades lose about a percent, and the whole system is paid for by the few that trend for months. The optional 3-day scratch (out if the close falls back inside the box within 3 days) is for drawdown-punished accounts: same engine, softer equity curve.

Cumulative sum of per-trade returns (1 unit per trade, entry-ordered). Gray band = 10th–90th pctile of random draws with identical gates, risk and exits. +2.25%/trade vs +1.36% random.

The entry’s edge survives every one of 16 exits (return pctile ≥ 98 in all). The 50 EMA trail extracts the most entry-specific alpha; the 3-day scratch trades some of it for a smaller drawdown. Fast ATR trails and fixed targets amputate the right tail that pays for this system.

The 5-bar box was inherited from the original lower-TF Bilbo Box study, so we tested it. The first sensitivity table made the full-episode box (box = the whole squeeze range, locked when compression ends) look clearly better — +3.35%/trade vs +2.25%. That number was wrong. “Compression ended today” is only knowable at that day’s close, but the flawed scan let the entry fill intraday on that same day, gated by the day’s own closing state. A resting order cannot do that, and most squeezes break out on exactly that day. The armable version — order can only rest from the following session, episodes that break on the resolution day are missed entirely — keeps just 384 of 1,295 trades:

The first-5 box wins on every metric once fills are honest, because its trigger is finalized mid-squeeze and a stop order can sit at the broker waiting for the resolution day — the day the episode box structurally cannot trade. The live spec is, and stays, the validated first-5 box. The invalid row is kept above deliberately: it was briefly promoted to the live config before the error was caught the same day, and this catalog doesn’t hide its mistakes.

Cumulative realized P&L (scr3 exit, 24-name live universe, booked at trade exit, from 2022-01), both variants armable-only. The first-5 box ends the window at +954pp vs +176pp for the armable episode box, which trades a quarter as often and lost money in 2022–23:

Why the flawed version looked so good: the trades it invented — intraday fills on the very day each squeeze resolved — are the best trades in the family. Remove them and the episode box keeps only late breakouts, a quarter of the flow. The general lesson, learned here for the price of a same-day config revert: any signal whose trigger level is finalized by the entry bar’s own close cannot be traded with a resting order, and backtests that pretend otherwise flatter exactly the trades that matter most.

Spec and read-bars fixed before the run (ret pctile ≥ 95, DD pctile ≤ 20). Universe rule fixed too: top 30 by 2-yr dollar volume, ex the 17. Shown: the deployment flavor (scr3). +0.47%/trade vs −0.04% random; both bars cleared by both exits.

Per-trade average by era — improving, not decaying:

High-beta names carry it; slow defensives are dead weight. Vol gradient is directional (high-vol half +1.46%/tr vs +0.72%) but not statistically settled at n=30.

The INTC-class trade — a squeeze that breaks out on the very day it resolves — can’t be captured by a resting daily order (the box isn’t final until that day’s close). The fix is the same one the hourly system uses: a fast execution layer. At each 5-minute close, today’s provisional daily bar is rebuilt and the daily compression flag is nowcast; a market buy fires the moment the squeeze is releasing and price crosses the running episode high. The resting first-5 order (K5) keeps priority — its trigger always sits at or below the nowcast’s — and the nowcast (NC) only ever trades the episodes K5 already died on. One trade per name per squeeze.

24-name live universe on intraday-derived daily data, 2003–2026: 1,914 trades, +1.51%/trade, PF 1.89, +2,888pp. On the 17-name discovery set the union added +42% more total P&L than the resting-order spec alone at equal risk-efficiency (sum/DD 24.9 vs 24.3) — the two paths fire on different squeezes by construction. The NC leg’s own showcase: INTC’s April-2026 squeeze resolution, +110% in 62 trading days, entered intraday at the 5-minute close that confirmed the release — with no knowledge of the daily close required. Both paths run in the live paper ledger from 2026-07-27; NC entries log a counterfactual scratch exit and a flicker tag so forward data settles the open design questions.

In the 2026 paper simulation, a UNH trade entered on a clean breakout (Jan 22) and was up comfortably — then UNH reported earnings before the open on Jan 27 and the stock opened 16% lower. The protective stop can’t help against an overnight gap: the exit fills where the market opens, not where the stop sits. A −7% worst case became −16.6%.

21% of all trades in the 24-name history held through at least one earnings report. Exiting at the close before every report and buying back the next day (only if the trend is still intact) beat holding on every measure — because mid-trend earnings gaps cut both ways about evenly, while the catastrophic ones only ever hurt the holder. The long runners survive the one-day pause: INTC’s +110% keeps +70 under the rule. Sleeve policy: flat by the close before every report, re-enter the next close if price still holds above the trailing stop.

Gross returns — no spread/swap/commission modeled; at 150–225bps/trade edge and 11–18 day holds the friction is real but secondary.

Survivor universe — both name lists are drawn in 2026; the defensible claim is beating random within the same names, which the baselines control for, not the absolute P&L.

No position cap — the book averages ~4 concurrent positions (max 13) at one unit each; a capped live version will differ.

Ranking noise — 16 exits on one dataset: the scr3-vs-ema50 sum/DD ordering flipped on the fresh names, exactly as the noise caveat predicted. Treat those two exits as equivalent; scr3 = lower absolute DD, ema50 = more P&L.

History spent — the 17 names’ full history was used in discovery; the 30-name run was the one-look validation. What remains is forward.