How a two-sided model beat a falling NKE on a razor-thin edge

NKE fell by half over the test window, yet the model made money on the way down by trading both directions, turning a falling stock into a positive result. This is a backtest of a rule-based NKE strategy, not a live forecast, and the live record is still too short to judge it.

The model standardizes a monitored NKE reading against its own recent history. When that reading drops to a statistical extreme it buys, expecting a snap back; when it spikes to the opposite extreme it sells short. In a falling tape the short side did the heavy lifting: the fastest and biggest winners in the book were shorts into overextended bounces. Across 88 trades the model was right 75% of the time and returned +144.55%, but the average trade earned only about 1.6%, so the edge is real but thin.

How the model is built, end to end.

The model watches one monitored NKE reading and measures how far it sits from its own recent average, in standard deviations. That single standardized number is the entire signal. A reading near its average keeps the model flat; only a stretch to a statistical extreme makes it act, and the direction of the stretch sets the direction of the trade.

An extreme low is read as oversold and opens a long; an extreme high is read as overbought and opens a short. Each position carries a protective stop and a profit target, but most trades end a third way: the reading drifts back through a calmer band and the model closes, having captured the reversion. The band sits inside the entry level on purpose, so the model does not flip on noise. In this falling tape the short trades, fired when the stock popped, were the cleanest and quickest wins.

The illustration follows the model's lineage on NKE end to end: from the single standardized reading, through the extreme-and-hysteresis rule, to the order that finally goes out. Each step is plain on its own; the discipline is in doing them in order, every time, on both sides.

At a glance

These figures come from the model's backtest measured under backtest conditions and against the registry's fixed acceptance gates. The total-return figure is the sum of the model's per-trade returns at a fixed position size, not a compounded equity curve; the buy-and-hold benchmark, by contrast, is a compounded hold of the stock, so the two are different measures.

Walk-forward verification

A trade walked through

On a steady afternoon for NKE, the model opened a long trade after several hours of patience. It waited through a shallow wobble in the wrong direction before the move it expected actually arrived. The position closed on a SIGNAL exit, with the broader market sitting in a mixed stretch, exactly the kind of regime the model was trained to read.

What the full trade record shows

Across its 88 NKE trades the model won 66 and lost 22. Almost every position closed on its own exit signal rather than at a hard barrier: 84 of the 88 exited on the signal, with just 2 take-profits, 1 stop-out, and 1 time-limit close. These are backtest results, not a forecast.

That is a two-sided, signal-driven system: it opens long or short on a stretched reading and almost always closes when the reading normalizes, not at a fixed target or stop.

The biggest winner was a short into a spike that closed at its target for +12.25%; a one-day short added +8.76%; the slowest winner was a long that gained +2.26% over 13 days.

The worst trade was a short that was squeezed against the position into its stop for -9.35%.

The edge here is thin and spread across many trades; no single position dominates the record, which is both a strength and the reason the per-trade return is small.

How does this compare to just holding NKE

Over the same window the model was tested on, simply buying NKE and doing nothing was its own kind of strategy. Comparing the two tells you whether the model was earning its keep or just riding the tide. In this case the model outperformed a plain hold, and the tiles below show the size of the 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 that were not used to build it. It is a consistency and replay-integrity test, not proof the model will make money live. A clean reproduction means the deployed rules behave like the studied rules; it says nothing about whether NKE will keep paying the same edge.

On NKE, the model is still quiet against its backtest distribution. Aligned means the model's live trades look statistically like the trades it took in training. Drifting means at least one statistic, hold time, hit rate, or exit type, has moved away from where the backtest sat, a sign the live record has started to diverge from the test. Quiet means there is not yet enough live evidence to call it either way. None of these labels prove the model good or bad; they describe how closely the live record is tracking the backtest so far.

When this approach fails

The model's failures are specific to mean reversion. The worst is a dip that is not a dip at all but a repricing: bad earnings or news resets NKE's fair value lower, the model buys the first leg down expecting a bounce, and there is none. The mirror risk is a short into a spike that keeps squeezing before it reverts, which is exactly how the deepest loss in the book happened. Volatility expanding beyond the band the model calibrated on widens both traps at once. None of these are bugs to engineer away; they are the standing cost of fading extremes in a trending stock.

Running this on NKE live, a desk watches three things. The first is drawdown, not just the current trade's drawdown, but the rolling drawdown of the strategy over the past several weeks, and a hard rule that pauses the model when it crosses a pre-agreed line. The second is liquidity and financing: thin books and high borrow or financing costs turn small edges into nothing. The third is alignment with the backtest distribution, when live trades start looking nothing like the backtest sample, the model is telling the desk something has changed.

Risk and honest limits

The model is checked against a fixed verification pipeline at each release: contract integrity, signal reproduction, trade parity, and behavior. Its current lifecycle state is shown below.

Lifecycle