How a condition z-score rule traded both sides of MDB's sharp swings in backtest

MDB went nowhere in a straight line over this window: it ran up, then sold off hard, then clawed back, ending roughly a quarter below where it started. A rule willing to work both directions traded those swings from both sides and came out ahead in backtest. What follows replays MDB's saved history through the model's fixed rules as a study of those rules, not a forecast or a recommendation, and with a double-digit single-trade drawdown at its worst and only a slim edge on the average trade, the record reads as backtest evidence rather than a settled edge.

The model watches a single MDB reading and measures how far it sits from its own recent range, waiting until that distance stretches to an unusual extreme before it does anything. While it holds no position, a quiet reading near its normal level opens nothing; only a real stretch opens a trade, going long on an upward extreme and short on a downward one. Once in a trade, the model stays in until the reading eases back toward its normal range, which is how almost every position closed. Over this window the gate opened 117 times, split 62 long and 55 short, and 65.81% of those trades came good for a running total of +214.88% at an even stake.

How the model is built, end to end

The reading the model is built on is one MDB market-state series, standardized against the level it itself usually shows so a move can be judged as ordinary or extreme by the stock's own history rather than an outside yardstick. A reading close to its normal band tells the model nothing and keeps it out. Only when the series strays far enough from that band, high or low, does the model treat the condition as met and a trade as possible, which is why long stretches of ordinary readings pass without a position.

A stretched reading is permission, not an instruction. Before committing, the model reads the direction of the stretch to pick its side, opening a long when the reading pushes to an upward extreme and a short when it pushes down, each with a protective floor and a target fixed ahead of time. From there the trade ends by one of a few routes, and on MDB almost every position closed the same way, when the reading eased back toward its normal range rather than by hitting a hard target or stop.

Picture the model laid across MDB from end to end: first the standardized reading, then the extreme it must reach before the gate opens, then the long or short that follows. Nothing in the chain is elaborate; its governing instinct is to stand aside until one reading strays far from its usual band and the direction of that stray is clear before a dollar is risked. The rule set's lineage rounds out the picture, for it was drawn from a wide field of machine-proposed candidates and held on through its backtest and walk-forward checks, a research-gated rule shown here on that backtest evidence rather than a proven live record.

At a glance

These figures come from a backtest of the model on MDB, scored against fixed acceptance gates, not from a live track record.

Walk-forward verification

A trade walked through

The walked example is a short held about three days. The reading stretched to a downward extreme with MDB near 343 dollars, the model sold, and the position ran to its profit target near 237 dollars for +30.85% in late May 2024. It is the model at its clearest: a short opened on a downward stretch and carried to its target, one half of a two-sided rule that also went long when the reading stretched the other way.

What the full trade record shows

Across its 117 MDB trades the model won 77 and lost 40. The exits leaned almost entirely on the reading easing back: 110 closed when the reading returned toward its normal range, 4 reached the profit target, 2 closed on a signal change, and 1 was stopped out.

A split this heavily weighted to the reading easing back is the mark of a model that closes when its condition clears rather than waiting for a fixed target, taking many small resolutions instead of a few large ones. That fits the thin per-trade edge: the average trade earned about 1.84%, so the record is built from volume, not from any one outsized win.

The biggest single winner was a short that ran to its profit target for +30.85% over about three days from near 343 dollars in late May 2024, the same trade walked above. Across the book the rule booked gains on both long and short trades, on MDB's up-legs and its down-legs alike, rather than leaning on either side alone.

The worst trade was a short stopped out for -15.30% within hours in April 2025 when MDB jumped back against the position near 145 dollars; that same trade is the deepest single-trade drawdown in the book, a 16.65% intra-trade adverse excursion, and it is one reason the drawdown gate is a bar the backtest missed.

No single trade carries the record, and at a 1.84% average edge the book depends on trading often and being right a little more than half the time. Because MDB round-tripped and ended lower over the window, the lesson is that a two-sided rule could come out ahead of simply holding by trading both the up-legs and the down-legs, but a 16.65% drawdown and a thin win rate keep it research-only.

How does this compare to just holding MDB

Over the same window the model was tested on, simply owning MDB would have ended about 26% lower, while the rule finished well ahead in backtest. Setting the two side by side is how you judge whether the rule earned its keep, and here it did, by trading both directions through a volatile round-trip rather than sitting through the full swing. The tiles below put numbers on that gap.

How well does the model reproduce its tape?

Walk-forward verification checks whether the saved rule path reproduces the expected signal behavior on held-out timestamps it was never fit on. It is a consistency and replay-integrity test, not proof the model will make money live. A clean reproduction means the deployed rules act like the studied ones; it says nothing about whether MDB will keep stretching this reading to the extremes the model traded across the test window.

In live trading the model has traded only once so far, and that lone read comes in misaligned with the backtest, far too little to draw any conclusion from. Treat live behavior as essentially unproven for now and lean on the backtest, read as exactly that, a study of how the rules behaved on saved MDB history rather than a live track record.

When this approach fails

The model's losses come from being caught on the wrong side of a snap. The reading stretches, the model takes its side, and MDB reverses hard before the move follows, which is how the deepest trade lost 15.30% on a short that was stopped out within hours in April 2025 as the stock jumped back against it. That same stop-out is the deepest single-trade drawdown in the book, a 16.65% intra-trade adverse excursion, and it is why the drawdown bar is one the backtest missed by a wide margin.

Three things are worth watching if this ever trades at size. The first is the drawdown, about 16.65% in the backtest, more than three times the acceptance bar, because a sharp reversal against a stretched reading can run well past the stop. The second is the thin per-trade edge, since the average trade earned only about 1.84%, below the bar the gates set, so the record leans on volume rather than a strong edge on any one trade. The third is the gap between live and backtested behavior, the first sign the reading is no longer stretching on the same moves it did across the test window.

Risk and honest limits

On this run the model's automated checks logged a caution rather than a clean pass, one more reason to weigh the rules on their record here rather than as a call on the stock.

Lifecycle