Strategy

This page documents the deployed configuration of the MAD-Velocity trading strategy — one book. The strategy is built on a single technical indicator, the Moving Average Distance (MAD), which measures the signed percentage distance between a stock's price and its own 20-day moving average. Standardized by each stock's trailing volatility, the reading sorts every trading day into one of six bands, from stretched below the average to stretched above it. Velocity is the indicator's five-day rate of change, split into the part produced by moving price and the part produced by the moving average itself. A stock is bought at the open after it jumps from the band below its average to the band above in a single close, with at least half of that move produced by price; candidates are drawn from the hundred most-traded names in the point-in-time S&P 500, and the book runs ten slots at ten percent of equity, filling the most-traded name first. A held name may be added to once, when it dips below its average while still climbing on price. When every slot is full, a new signal may be admitted on borrowed money to a hard cap of 130% of equity, financed at the federal funds rate plus 1.5%. A position is sold at the open after its stretch rolls back — a close between one and two standard deviations above the average followed by a close back inside one. Sessions settle in a fixed order — sells, forced de-leveraging, adds, then buys. Simulated over 2016–2025 from $100,000 with five basis points of cost per side, the book finished at $1,060,457 — 26.7% annualized, Sharpe 1.0. The same configuration trades this site's book forward, marked to market daily.

I. Introduction

MAD-Velocity is the product of an open quant research program conducted in four published papers. The first [1] established that a stock's position within its displacement band forecasts its next regime transition. The second [2] converted that forecast into an entry gate by decomposing the change in displacement into its price and average components. The third [3] showed the joint displacement–velocity dynamics to be stationary and shared across the index. The fourth [4] built the portfolio the signal implies and found that nearly everything of consequence happens in the step between signal and book. The gated entry fires roughly twelve times per trading day across the index, far more than a concentrated book can act on, so realized performance is decided by which candidacies receive capital. The selector that survived every test is traded volume.

Publication did not end the testing. A registered program ran the published book against alternative selectors, capacity configurations, and overlays — fifty-five runs whose complete registry, scripts, and outputs are public in the MAD-Portfolio repository [4], reproducible from fixed seeds. Out of that program came the two mechanics the deployed book carries beyond the published rules, an add into held weakness and a margin bridge for overflow signals, and the combined configuration this page documents — one book, simulated end to end over the same decade as the published work.

The sections that follow state the configuration precisely and present its record. Section II defines the indicator. Section III states the trading rules. Section IV describes the data and the simulation protocol. Section V presents the results, component by component and combined. Section VI describes the live book that trades the configuration forward on this site, and Section VII concludes.

II. The Indicator

A. Construction

The workhorse of technical analysis is the moving average, and its standard use compresses the price path into crossover events — price above or below its average, a fast average crossing a slow one. What that compression discards is the quantity this program studies — how far price sits from its own average, and how that distance moves. Let P_t denote a stock's closing price. The trend reference is the 20-day simple moving average,

Fig. 1.  SMCI, daily closes and the 20-day simple moving average, February–August 2026. The average trails the price and smooths it; the gap between the two lines at any close is the raw material of the indicator.

The Moving Average Distance is the signed percentage displacement of price from that reference,

The series is positive when price trades above its average and negative when below. A reading of +4 means the price closed four percent above its average.

Fig. 2.  The window of Fig. 1 rendered as MAD — the two lines collapse into one series, the gap as a percentage of price. The March collapse reads −51% because price fell while the average still carried the old level; the June run peaks at +28% with price far ahead of its average.

A fixed displacement is not equally meaningful in every name or every market — a four percent stretch is an ordinary day in a volatile stock and an event in a quiet one — so MAD is standardized by its own trailing variability, the sample standard deviation over a window of 255 trading days,

and the standardized displacement

expresses displacement in units of that stock's own typical stretch, comparable across time and across names. The standardization lets a single rule serve the entire cross-section. Every quantity in (1)–(4) is strictly trailing, so a signal computed at one day's close is executable at the next day's open with no revision and no look-ahead.

B. Regime Classification

The standardized series is discretized into six ordered regimes at thresholds ±1 and ±2, with no neutral state — every observation carries both a direction and a magnitude. Where a stock sits in these bands, and how it moves between them, is what the strategy watches [1].

Fig. 3.  The standardized series for the same window, each close colored by the regime band it lands in. All six regimes are visited — the late-March closes sit deeply oversold below −2, the run crests extended above trend at the turn of June, and the tape spends most of its days in the two bands nearest the average.

Table I. Regime definitions.

C. The Velocity Decomposition

A regime transition is, mechanically, a change in MAD — and that change is not a single thing. MAD can rise for two very different reasons — price rising toward its average, or the average falling toward price as old highs leave the 20-day window. Writing (2) as MAD = 100 · (1 − SMA/P), a first-order expansion of a change over k days separates it into a price term and a moving-average term,

with k = 5 trading days throughout. The first term is positive when price rises — displacement improving because the security is genuinely recovering. The second is positive when the moving average falls, typically because old highs are leaving the window, so displacement can "improve" while price is flat or still declining. The signed share of the change attributable to price is

where the numerator is the price term of (5) and the denominator sums the magnitudes of both terms. The share lies in [−1, +1] — values near +1 mark a price-dominated move, values near −1 a move manufactured by the average alone. This decomposition is the velocity of MAD-Velocity, and it functions not as a forecaster but as a qualifier, identifying which regime transitions are real [2].

Fig. 4.  The five-day change in SMCI's MAD split into its two components. On July 22 the move is price-dominated — price term +10.9 against average term +4.5, s = 0.71 — the genuine transition the entry gate accepts. On April 6 the same-sized "improvement" is manufactured — average term +10.7 against price term +0.4, s = 0.04 — the crash highs leaving the 20-day window while price sat still. The March plunge in the price term shows the collapse itself was genuine price movement.

III. The Trading Rules

A. The Entry

A buy signal is a regime transition with a credential. The transition is the cross — the stock closes in the −1 band one day and in the +1 band the next, price passing from just below its average to just above it in a single close. The credential is the velocity gate — the five-day share of (6) must show at least half the move coming from price, s ≥ 0.50, so that transitions manufactured by a falling average are refused [2]. A signal that passes both tests is good for exactly one day. The stock is bought at the next session's open, or the candidacy expires untaken.

B. The Exit

The exit is the rollback. After a position closes above trend — the +2 band, more than one standard deviation stretched — the first close back in the +1 band sells the position at the next open. There is no stop-loss and no price target. A position is held, through gains and losses, until the rollback ends it.

Fig. 5.  The rules on the tape. Seven crosses fired in the window; the gate passed four and refused three — among them April 10, where s = 0.49 missed the bar by a hundredth, and July 30, where the cross was almost entirely the average's doing (s = −0.92). Five rollbacks mark the exits, including June 5, two sessions after the run's crest. The July 22 gated entry still shows open on the signal ledger.

C. The Book

The universe is the point-in-time S&P 500 — every name that was a member on a given day, including the members that later vanished. At the final session before each month boundary, the constituents are ranked by shares traded; the top hundred form the bench, and the bench is frozen until the next month begins. Only bench names may be bought. The account runs ten position slots, each taking ten percent of equity at entry, and when more gated signals arrive than open slots remain, the most-traded name fills first [4].

D. The Adds

A held name may be added to once. The add fires when the stock closes below its average with its displacement improving and the improvement led by price — in the terms of Section II, z < 0, the five-day change in MAD positive, and s ≥ 0.50. The add takes a full slot of its own at ten percent of equity.

E. The Bridge

When the book cannot fund an admission from cash — a gated signal or an add arriving with every slot deployed — the position may be financed on borrowed money. Gross exposure is capped at 130% of equity, and the cap is hard. Borrowing is charged daily at the effective federal funds rate plus 1.5%. If losses push gross exposure past the cap, the newest positions are sold first at the next open until the account is back inside it.

F. Session Order

Every session settles in the same order — sells first, then any forced de-leveraging, then adds, then new buys. Nothing in the configuration asks for judgment at the close; every rule above is mechanical.

IV. Data and Simulation Protocol

A. Data

The price record is daily bars from polygon.io, split-adjusted and dividend-free, and every series on this page — the strategy, the index controls, the random books — is measured on that identical basis. Membership is point-in-time — the S&P 500 as it stood on each trading day, taken from the fja05680 membership history. Across 2016–2025 the universe contains 728 distinct members, roughly five hundred on any given day, and the members that later vanished — acquired, delisted, or collapsed — remain in the test on the days they were present. A backtest that quietly drops the casualties inherits their absence as alpha; this one carries them. The data vintage is pinned at July 16, 2026, and the runs of Section V record SHA-256 fingerprints of their inputs, so the exact tape behind every number is identified.

B. Simulation Mechanics

The account starts at $100,000. Execution follows the conventions of Section III — quantities computed at one day's close act at the next day's open, and no rule reads anything the close had not already printed. Each side of every trade pays five basis points. Slots are sized at ten percent of equity at entry, and sessions settle sells, forced de-leveraging, adds, and buys in that order. The simulation runs the full decade 2016–2025, and following [4] the record splits at 2022 — a configuration is judged on its 2022–2025 confirmation window as much as its full decade, and a rule that worked only in 2016–2021 fails the program's standard.

C. Protocol and Reproducibility

Every number in Section V descends from registered runs with fixed seeds. The component studies were pre-registered against explicit null models before their code ran, and the complete registry — fifty-five runs, each with its null standing and verdict — ships with the scripts and outputs in the MAD-Portfolio repository [4]. Reproduction is enforced rather than promised: each study begins by re-deriving the published baseline and hard-exits unless it lands on $552,887.57, and the combined run of Section V refuses to print unless all three of its component results reproduce to the cent. A reader with the repository and a data key can rebuild the store and every table on this page.

V. Results

Section IV fixed the tape and the mechanics. This section prints the record — first the picture, then the numbers, then each component on its own terms.

Fig. 6.  The decade, four ways. The published baseline, the adds alone, the bridge alone, and the combined book, with the SPY and RSP controls dashed beneath them — identical tape, identical costs, an identical dividend-free basis under every line. Every curve starts at $100,000 on the first session of 2016.

Table II. The record, 2016–2025. Full-decade and confirmation-window (2022–2025) statistics; all series dividend-free on an identical basis.

A. The Baseline

The published rules alone — the bench, the ten slots, the volume fill — finished the decade at $552,887.57, 18.7% annualized against SPY's 13.1% and RSP's 9.8%, at a Sharpe of 0.90 [4]. Its confirmation window was stronger than its full record — 19.7% annualized, Sharpe 0.97, a worst drawdown of −19.1% against SPY's −25.4% — and it was the only tested configuration with that property. Against two hundred random-selection books run under identical mechanics it beat every one; the random median finished at $242,100.

B. The Adds

Adding once into held weakness lifted the decade to $754,818.05 and the Sharpe to 1.04, on fewer trades — 583 against the baseline's 721, with 145 adds taken. Its registered settlement reached the 94.4th percentile of a 500-seed random-add null, short of the 95th-percentile bar.

C. The Bridge

The bridge alone lifted the decade to $971,305.16 — 25.6% annualized — at the cost of the deepest drawdown on the page, −49.6%. It admitted 870 positions on borrowed money, paid $37,409 of interest, and was forced to de-leverage nine times. It cleared its full-window null at the 96.6th percentile and failed the confirmation window under the registered forced-liquidation remedy.

D. One Book

Wired together, the full configuration finished the decade at $1,060,456.71 — 26.7% annualized, Sharpe 1.0, worst drawdown −40.6%. The run reproduced each component's final value to the cent before printing, and its ledger reads 741 trades at a 74% win rate, 214 adds of which 201 drew margin, 669 bridge admissions, 18 forced de-leveragings, and $42,704 of interest paid. Its confirmation window came in at 19.0% annualized against the baseline's 19.7% — the extensions' work is concentrated in the full decade. The components' registered standings are stated above and do not change by being combined; the combined run is the deployed wiring, simulated end to end.

VI. The Live Book

Simulation ends where this site begins. The configuration of Section III trades forward in a paper account, marked to market daily, with every output public as it happens. Signals are computed at each close and published on the Signals page before their outcomes are known. The Book holds the open positions, the Archive holds every completed round trip, and Returns keeps the week-by-week score. None of it can be revised after the fact.

The forward record opened in August 2026. It is short, and it will stay short for a while — that is the nature of the only evidence that accrues one session at a time. For the two mechanics that missed their registered bars in simulation, this ledger is the test that settles them, forward and in public, on tape that did not exist when they were designed. The page will not be rewritten to fit the outcome; the outcome will simply arrive.

VII. Conclusion

This page set out to document one book — the deployed configuration of the MAD-Velocity strategy — from its indicator to its record. The indicator is a single number per stock per close. The rules that trade it are mechanical throughout. The record is a decade simulated on a pinned tape, $100,000 to $1,060,457, reproducing its registered components to the cent. A decade of testing reduces to one sentence — the signal says when, volume says which, and nothing run against that pair has beaten it. The published rules remain the validated result; the extensions that trade beside them are documented with their standings stated. What remains is the only test that counts, and it is underway on this site.

References

1. Arrington, L. (2026). Position Within Trend: The MAD-Markov Model for Calibrated Regime-Transition Forecasting. github.com/madvelocity/MAD-Markov-Model
2. Arrington, L. (2026). MAD-Velocity: Constructing and Evaluating a Moving-Average-Distance Trading Signal. github.com/madvelocity/MAD-Velocity-Signal
3. Arrington, L. (2026). MAD-Manifold: Mapping and Testing the Phase-Space Flow of Moving-Average Distance. github.com/madvelocity/MAD-Manifold
4. Arrington, L. (2026). MAD-Portfolio: Portfolio Construction from the Moving-Average-Distance Signal. github.com/madvelocity/MAD-Portfolio