Methodology
Source & status
The rules behind Dipbot come from the @MrMilkTrading thread of 7 September 2026. The underlying strategy was candidate 18 out of 465 overnight backtests, and the author still labels the results UNVERIFIED. The rules were frozen that day and this site only paper-tracks them — no orders, no exchange keys, no discretion. The full trade list and the reproduction code are public at original strategy research.
Universe
Twenty Binance USDT-margined perpetual futures — 1000PEPE, AAVE, ADA, AVAX, BCH, BNB, BTC, DOGE, ETH, FIL, LINK, LTC, NEAR, SOL, SUI, TRX, UNI, WLD, XRP and ZEC. Daily bars, UTC days. The universe was selected in August 2026 and applied to every window, which is a known survivorship bias.
The signal — evaluated at each daily close, per coin
Peers and eligibility
A coin's peers are the other nineteen. A peer counts only if it has a complete five-day return both today and yesterday, and at least sixteen of the nineteen must qualify. Both shares use that same eligible set, so a new listing or a data gap cannot manufacture a transition.
Breadth share
The positive share is the fraction of eligible peers whose five-day close-to-close return is above zero. Peers with exactly zero return count in the denominator only.
Long trigger
Today's share is 70% or more, yesterday's was below 70%, the coin's own five-day return is positive, and today's close is above yesterday's. Short is the exact mirror on the negative share.
Stop distance — daily ATR
2.5 × ATR(14) on the daily timeframe: fourteen daily bars, simple mean of true range — not Wilder's smoothing, and never 4H or 1H. It is measured at the daily close, expressed as a fraction of that close and floored at 0.5%. If it would be wider than 20%, the signal is skipped entirely.
The trade
Entry
At the open of the day after the signal. In the minute engine that is 00:01 UTC; the daily package uses the daily open.
Stop and target
Stop at entry minus (long) or plus (short) the daily-ATR stop distance. Target at three times that distance — a 3R target. The research engine watches both on minute bars after entry; this daily package checks them on the daily bar, stop before target.
Time exit
Anything still open on the fourteenth day exits at that day's open.
One slot per coin
A signal in a coin that already has an open position is skipped, not stacked.
Sizing and the account
The paper account starts at $100,000 and is unlevered. Each eligible signal wants 0.25% of open-time equity measured to its stop, including a 20 bp cost reserve. All signals admitted on the same day form one batch and are scaled pro-rata so open initial risk never exceeds 1.5% of equity and gross notional never exceeds 2× equity. The cap is enforced at admission, not continuously. Quantity is floored to 1e-8 and orders below $100 notional are not placed. So $250 is the maximum single-signal risk on the starting account, not the typical one — in the research run the median reserved risk was $104 and the median filled position $777.
Cost model
20 bp round trip — 10 bp per side, being 9 bp fee plus 1 bp adverse slippage — debited at entry and at exit, plus the actual 8-hour funding settled while the position is held. Funding can be a debit or a credit. The reproduction package also ships 12 bp and 30 bp scenarios.
How the engine processes a day
Order of operations matters, and getting it wrong leaks future information into sizing. Each day the engine settles funding first, then processes known opening-gap exits, then sizes and admits new orders using open-time equity with one slot per coin and pro-rata batch scaling, then applies intraday protection, and only then marks positions at the close. Inside a single bar, the stop is checked before the target — the conservative assumption — and a target requires strict penetration. Known exits are always processed before new entries.
What this page is not
There is no order execution here and no exchange trading key: this is a paper track only. Forward evidence started on 2026-09-07, so the forward sample is tiny and every forward ratio should be read as noise until it is not. The backtest section shows the deflated Sharpe under three different trial counts, the Monte Carlo drawdown distribution, and the checks that have not been run.