RLTW v3
Ride Like The Wind
The Nasdaq model: universe aligned with MomentumCommand’s Nasdaq-100 ranking (~100 names), regime-switch (cash in bear), Top 8 equal-weight at entry, 20-day min-hold + hysteresis, and let-winners-run (strong names may grow to 25%, then trim to 20%). It is not Core. RLTW can run a looser Buffett gate (quality > 0.25) because it already fishes in liquid Nasdaq large-caps; Core, on the whole market, requires 0.35.
Origin and inspiration
The architecture stems from observing MomentumCommand (momentumcommand.com): the Top 10 Nasdaq-100 names by momentum in a “Risk On” regime, switching to Treasury bills (BIL) when the regime deteriorates, weekly execution. Its stated returns must be read with caution: simulations, excluding commissions, with survivorship bias. We adopted its architecture and theoretical roots:
- Dual-Momentum (Gary Antonacci, 2014): absolute momentum as a regime switch (out of the market in a downtrend, to cut left-tail risk) and relative momentum to select the strongest names.
- Momentum as a factor anomaly (Jegadeesh-Titman): relative momentum over medium horizons is one of the most documented risk premia.
Why the RLTW ranking does not match MomentumCommand
We share the architecture (Nasdaq, Top N, cash in a bear), not the ranking rule. MomentumCommand ranks by price return already earned since Stage 2 / Risk On began (e.g. from 10 April 2026): the names that have already risen the most sit on top. RLTW ranks by buy_priority: the ML probability that the name is a relative winner vs QQQ over the next ~90 days, times remaining upside, penalised if the rally is already crowded (anti-chase). “Mom 65%” does not mean “the chart is going up”: a pullback from the highs can stay in the Top 8, while a bull leader (+100% since April) can be blocked by the Buffett gate or the reverse gate. Ranking data is yesterday’s close (prices at 06:30, features and model at 07:15, Italy time). The 0–100 SELL score on the ticker page is the Buffett screener, not the RLTW score. A 50/50 hybrid rank (ML + return since Risk On) was tested and rejected: same return, nearly double the drawdown.
See the live ranking →Why RLTW is built this way
Nasdaq only, on purpose
The model’s predictive momentum is stronger on growth/tech (AUC ~0.61–0.63) and weaker on value/cyclicals (~0.54). RLTW accepts that evidence: instead of diluting the signal across the whole market, it applies it where it has an edge.
Looser Buffett gate (0.25 vs Core 0.35)
Whitelist names are already liquid large-caps, often with cash and a moat. A 0.35 bar would cut too many growth leaders (high multiples) without adding protection. 0.25 is there to drop pathological cases (debt, no FCF), not to run strict value investing.
Eight positions, 12.5% at entry
On a narrow universe conviction is expressed as concentration, not as 15 slices. At purchase each leader is ~12.5%. Winners may then run: we do not rebalance every week back to 12.5%. Only if a name exceeds 25% of equity is it trimmed to 20% (let-winners-run, RLTW v3).
Cash in a bear, not “hold on”
Momentum crashes in drawdowns are documented. On a 100% Nasdaq book, staying invested in a bear is the risk to avoid. Core, being multi-sector, does not need this switch.
The RLTW 6-step pipeline (nasdaq_apex profile only)
All 6 steps are live in production (RLTW v3): QQQ regime, Nasdaq-100 whitelist (MC ranking), Buffett gate 0.25, ML ranking, Top 8 at 12.5% entry with 25%→20% trim, 20-day min-hold + rank ≤ 14 hysteresis.
Regime Detection (macro on QQQ, rules-based)
In productionBull / Neutral → full regime, Top 8. Early Warning → high-conviction only (MomProb ≥ 52%). Bear → 100% cash / Treasury and full exit from positions. Why: Dual-Momentum says leave the market when the lead index is in a downtrend, before momentum crashes do the damage.
Nasdaq Universe (whitelist)
In productionMomentumCommand’s Nasdaq-100 ranking (~100 names), plus holdings and live lots. Updated from the scrape: a new name on their ranking enters on its own. Excludes NYSE, European listings and non-tech by RLTW design. Why: the ML edge is concentrated in Nasdaq growth, not the global market.
Quality & Value Gate (Buffett)
In productionExcludes names with insufficient fundamental quality (fundamental_quality < 0.25). Looser than Core (0.35) because the universe is already Nasdaq large-cap: the filter drops pathological cases, not high-multiple growth.
Predictive Momentum Ranking (XGBoost + entry filters)
In productionRanking on buy_priority = ML probability × remaining upside × anti-chase. Regime-aware reverse gate (Rev ≤ 0.55 / 0.60 / 0.70 on RLTW bull). Why: in a confirmed bull momentum persists; allowing a bit more chase is consistent with the literature, not FOMO.
Concentrated Position Sizing (equal-weight + let winners run)
In productionTop 8 by buy_priority; ~12.5% at entry → ~100% invested in bull. Winners are not rebalanced to 12.5%: they may run up to 25% of equity, then trim to 20%. New buys are sized on current equity (not initial capital). Cash remains only if fewer than 8 candidates pass the filters, or after a trim. Why: letting leaders run is where most of the alpha sits; weekly 12.5% rebalancing would cut the names that work.
Position Management (20d min-hold + hysteresis)
In productionFull exits: BEAR → sell all; holding < 20 sessions → hold (even if MomProb dipped: check-up says “hold tight”, not “sell”); after 20 days, rank ≤ 14 → hold; rank > 14 → sell. The 25%→20% trim is a partial “Reduce”, including during min-hold (a risk cap, not an exit). Check-up aligned with the Wizard. Why: without hysteresis the book rotated too often (~221 round-trips); M2 cuts turnover by 35%.
How we validate RLTW (plain language)
A model does not “work” because the chart looks good. It works if, after costs, it beats a lazy honest choice: buy the Nasdaq index and do nothing. That index is QQQ. Two ideas.
1. The QQQ comparison
RLTW only buys Nasdaq names. The fair test is: “if instead of following it I had bought QQQ over the same period, with the same money, who would have won?” The difference (after commissions) is alpha. Positive alpha = RLTW added value. Negative alpha = you would have been better on the index.
Same period, same rules
We replay history day by day: RLTW and QQQ start together on the same calendar. We do not pick the most convenient slice of history.
No peeking at the future
The model is trained only on data that had already happened, then asked about later days (walk-forward). Like studying an old textbook and sitting an exam on a new chapter: using the exam to study is cheating.
Costs are paid
Every buy/sell costs (commissions, and for an EU investor also EUR/USD conversion). A model that “beats QQQ” before costs can lose after them. That is why we always show the net scenario.
One number is not enough
We look at return, alpha vs QQQ, how painful the worst drop is (drawdown), and how often the book is rotated. Extra return from spinning names every three days is often an illusion: commissions eat it.
2. What happens after every change
When we change the ML brain or an RLTW rule (holding, hysteresis, filters), the novelty does not go live by eye. It passes two gates. If it fails, the old model stays.
Gate A — the ML brain (every week and after every retrain)
The new model trains in a separate room (it does not touch the live one). It must beat chance on unseen data (momentum AUC at least 0.53; reverse at least 0.50) and not drop more than 0.02 versus production. If it passes: automatic backup of the old one, then promotion. If it fails: production unchanged. A report is written.
Gate B — RLTW rules (one change at a time)
One novelty at a time, compared with previous RLTW and with QQQ, at the same costs. To promote it we need: positive alpha vs QQQ over 1 year and 2 years, fewer portfolio rotations, drawdown not worse, and results that do not collapse if you nudge the parameters. QQQ is the yardstick, not a target to “force” by retuning until RLTW > QQQ pops out.
| Exam | When | What must happen | If it fails |
|---|---|---|---|
| ML brain | Weekly retrain (Sunday) and after every training change | Holdout AUC above the floor and not worse than live | Keep the old model + backup |
| RLTW rules | After every pipeline change (e.g. 20d hold, hysteresis) | Better or more efficient than current RLTW, and alpha vs QQQ > 0 on 1Y and 2Y | The new rule does not go to production |
⚠️ Honesty: these tests use a single market window (mostly-bull Nasdaq) and a whitelist of names that exist today (survivorship). A past result vs QQQ is not a promise about the future. The point is precisely not to fall in love with a chart.
RLTW vs Core profiles
Same ML, two jobs. Core covers the whole market with a tighter Buffett gate. RLTW is Nasdaq-only, more concentrated, with a looser Buffett threshold because the universe is already filtered:
| Dimension | Core profiles | RLTW v3 |
|---|---|---|
| Universe | Broad multi-sector (load_universe_tickers) | Nasdaq-100 (MC ranking, ~100 names) |
| Buffett gate | Quality > 0.35 | Quality > 0.25 (looser) |
| Positions | 8–15, priority-scaled sizing | Top 8: 12.5% at entry; 25%→20% trim |
| Bear regime | Does not force cash | 100% cash + sell all |
| Entry | Conservative Mom ≥55% · 25% sector cap | Mom ≥42% on Nasdaq · no sector cap |
| Exits | sell_priority | 20d min-hold + rank≤14 hysteresis + v3 trim |
| Calendar | Wizard when needed | Every market day (morning mail). EMC/MC: Friday→Monday only |
The selection engine
Selection is an XGBoost model whose target is the cross-sectional “90-day relative winner” — the probability that the stock lands in the top 40% by outperformance vs the market over the next ~3 months, the horizon on which the momentum anomaly becomes learnable.
Holdout AUC ~0.636 on the Nasdaq. Modest but real edge, concentrated in growth: on value/cyclical names AUC drops to ~0.54. RLTW v3’s value is the combination of selection + regime-switch (cash in bear) + concentration at entry + let-winners-run + exit management (M2).
How we stand vs MomentumCommand
Do not add their YTD to our walk-forward: they are two different exams. The honest comparison is the table below: same window from 2 January 2025, same simulator, 0.19% fees even on the official replica of their rules (no longer the 13-week proxy).
| Book | €10,000 becomes | vs QQQ | Sharpe | Worst drop |
|---|---|---|---|---|
| QQQ (buy and hold) | ~€13,890 (+39%) | yardstick | — | — |
| Official MC (Top 10, Stage 2, 0.19%) | ~€41,720 (+317%) | +276 points | 1.9 | −40% |
| EMC (Top 8, 40% × 4w blend, 0.19%) | ~€51,120 (+411%) | +372 points | 2.1 | −37% |
| RLTW v3, Nasdaq-100, 0.19% | ~€33,300 (+233%) | +194 points | 2.8 | −23% |
Different commitment: MC and EMC take one decision a week (Friday close → Monday open, Friday-night mail). RLTW may buy and sell any market day (morning mail, 20-day min-hold). MC is the official replica (Top 10). EMC is the same DNA with 8 names and 60% Stage 2 + 40% last-4-week rank. At 0 fees MC is +331% vs the site ~+336%. In this bull EMC made more; RLTW drew down less (−23%). Survivorship, mostly-bull window.
What MomentumCommand publishes (their tape, 0 fees)
Their figures, refreshed from our scrape. Calendar 2026, commissions excluded. They explain the public claim, not the score vs RLTW.
| Window | MomentumCommand | QQQ |
|---|---|---|
| YTD | — | — |
| 6 months | — | — |
| Month | — | — |
| Sharpe | — | — |
| Max drawdown | — | — |
Waiting for scrape…
The MC site (0 fees) from early 2025 is ~+336% and YTD ~+129%. Our 0-fee replica on the same tape returns +331% and YTD +118% (6-month both +38%). With 0.19% fees the replica is +317% / DD −40%; RLTW v3 is +233% / DD −23%. Do not add the two YTDs.
The three live books (same capital, since we switched them on)
USD 10,000, same closes. The % is vs those 10,000 (0.19%/side already in NAV). It is early: it shows who rotates from here, not who “won” 2026.
Live RLTW vs MC vs QQQ curve →A concrete example: RLTW v3 vs “just buy QQQ and sit”
Walk-forward over about 20 months, full Nasdaq-100 universe (MomentumCommand ranking), today’s RLTW v3: holds at least 20 days, does not sell at the first slip, lets winners run to 25%. Question: €10,000 in RLTW or €10,000 in QQQ, same period? QQQ made +38.9% (€10,000 becomes ~€13,890). Two repeats, same outcome.
| What you did | Your €10,000 becomes | Versus QQQ | Smoothness (Sharpe) | Worst drop |
|---|---|---|---|---|
| Buy QQQ only (the lazy choice) | ~€13,890 (+39%) | the yardstick | — | — |
| RLTW v3, Nasdaq-100, 0.19%/side broker | ~€33,300 (+233%) | about +194 points: you beat QQQ | 2.8 (good) | −23% |
One line: with the universe aligned to MC’s ranking, RLTW v3 stays well above QQQ after broker commissions. This is not MomentumCommand’s YTD (different window, they run at 0 fees). The cost is a wider worst drop than when we measured ~56 names (−23% vs −11% then). 0.29% (EUR/USD) has not been re-run on this universe: we do not invent it. Survivorship, a single bull window: not a promise about the future.
The model’s history (M0 → v3)
Each version is one change, measured against the previous one and against QQQ. The table above is RLTW v3 on the 100-name universe, current window (QQQ +39%). The M2/v3 “56 names” rows below are the 25 August window (QQQ +36%): do not add them or compare 1:1 with +233%.
| M0 — old version (rotate every ~10 days) | ~221 cycles · hold ~18 days. At 0.29% it lost to QQQ by a hair |
| M1 — hold 20 days | ~166 cycles · hold ~24 days |
| M2 — 20 days + don’t sell on the first slip (from 23 Aug 2026) | ~143 cycles · hold ~28 days. 25 Aug window: +97% at 0.19% / +56% at 0.29% |
| v3 — let winners run (trim 25%→20%), ~56-name universe | 25 Aug window: +140% at 0.19% / +64% at 0.29%. Worst drop −11% / −6% |
| v3 — in production today: same book, full Nasdaq-100 | 28 Aug window, 0.19%: +233% vs QQQ +39%, Sharpe 2.8, worst drop −23%. 185 trades, 55 names touched |
| Challenge — same APEX Top 8 on every DB ticker | Rejected: +101% vs +202% for Nasdaq APEX in the same process. It did buy energy, industrials and materials, but alpha was roughly cut in half (+62 vs +163 points) and turnover rose. Nasdaq remains a special universe for this model. |
| How many names in the book | 8 max; 12.5% at entry, up to 25% if they win, then trim to 20% |
⚠️ Letting winners run raises return and also drawdown. Fewer round-trips help you most if you pay EUR/USD conversion. Past results do not guarantee the future.
Bear-market prediction & front-running
Idea: reactive algorithms sell all at once when QQQ breaks the SMA200; whoever anticipates it sells at better prices. Today the regime is detected with a rules-based score (QQQ trend, VIX, Nasdaq breadth) — not an ML model. Stages emitted:
- BULL_RISK_ON — full operating regime on the Top 8.
- NEUTRAL — selective operation.
- EARLY_WARNING_DERISKING — raised entry threshold (MomProb ≥ 52%), trimming into rallies.
- BEAR_RISK_OFF — 100% cash / Treasury.
⚠️ Why not an ML bear model? Tested: AUC ~0.42 on ~45 events in ~2 years (data-starved). It needs more history and macro features (yield curve, credit spreads, VIX term structure). Until then the rules-based regime is more robust. Future work, not a current capability.
Glossary of the metrics
⚠️ Notes and limits
- The extraordinary returns cited for similar models are simulations with survivorship bias and excluded costs.
- The selection edge is real but modest and concentrated in growth (AUC ~0.64 Nasdaq, ~0.54 non-Nasdaq).
- Regime detection is partly reactive (SMA 200 / trend): it moves to cash after an initial 5-8% drop.
- RLTW v3 raises return by letting winners run; on the 100-name universe the simulated worst drop is ~−23% at 0.19% (it was ~−11% on the 56-name measure).
- AI Stock Mover does not send orders to a broker: it provides the plan, execution stays with the user.
Scientific references
Every model choice rests on the quantitative literature. Key sources:
- Jegadeesh & Titman (1993), Journal of Finance — Cross-sectional momentum (buy winners, sell losers) over 3-12 months: basis of the 'relative winner' target.
- Carhart (1997), Journal of Finance — Momentum as a fourth return factor (UMD): confirms its empirical robustness.
- Moskowitz, Ooi & Pedersen (2012), J. Financial Economics — Time Series Momentum — In a confirmed trend momentum persists: justifies relaxing the reverse gate and chasing in BULL regime.
- Asness, Moskowitz & Pedersen (2013), J. of Finance — Value and Momentum Everywhere — Momentum and value work across markets/assets: foundation of the value-gate + momentum combination.
- Antonacci (2014), Dual Momentum Investing — Absolute momentum (go to cash when the index is weak) + relative momentum: RLTW's Regime-Switch architecture.
- Daniel & Moskowitz (2016), J. Financial Economics — Momentum Crashes — Momentum crashes in bear markets/rebounds: justifies going 100% cash in BEAR regime.
- Barroso & Santa-Clara (2015), J. Financial Economics — Momentum Has Its Moments — Risk-managed momentum (volatility-scaled): justifies the protections and sizing discipline.
- Bernard & Thomas (1989/1990), J. Accounting Research / JAE — Post-Earnings-Announcement Drift — Post-earnings drift lasts weeks/months: consistent with the ~90-day signal horizon.
- Jegadeesh (1990) & Lehmann (1990) — Short-term reversal — At 1 week/1 month reversal dominates (not momentum): why the old 10-day target was noise.