Where the app stands today — read this first

Almost every advised book comes back 100% cash — 32 of the 35 built on 14 August. That is the correct output of an honest bar, not a malfunction and not a bug to fix by lowering it.

A walk-forward score no longer faces one fixed threshold. It must beat its own out-of-sample standard error — which is roughly 1/√(OOS years), so a 1.5-year score is judged far more harshly than a 12-year one — multiplied by the expected best of N pure-noise trials. The swing stocks build priced its candidates against 2,808 lifetime trials and produced zero positions; across the other scopes and horizons that day the trial counts ran from 627 to 46,413.

The three books that were not empty are worth looking at closely, because they show you exactly what the bar is for. They only filled at the lowest quality setting, where the deflated bar itself governs (≈0.48–0.60); push the quality slider to 1.0 and it binds instead, and they empty too. Of the five positions in the indices book, three were eToro index CFDs picked by the same calendar rule (dow_effect) on 997–998 bars — the eToro candle feed's 1,000-bar ceiling — 1.5 out-of-sample years, and three windows each. That is the precise cohort this bar was calibrated to distrust: a short feed makes the standard error large, and three windows is barely a measurement. Only two positions rested on real history (^NDX and ^STOXX50E, 12.5 OOS years over 25 windows) — and they scraped through at 0.61 and 0.55.

So a book with a few names in it is not a green light. Open each position and read oos_years and the window count before you believe it.

The bar it replaced was 0.30 + 0.1·log10(breadth) — three constants with no derivation, and the same number for a 1.5-year score and a 12-year one. It admitted about a fifth of a pool whose spread of scores was statistically indistinguishable from pure noise, and it handed scores resting on four windows or fewer roughly six times the qualification rate — the cohort whose mean out-of-sample Sharpe was the worst of any.

So read a cash-heavy book as what it says: on this evidence, at these costs, almost nothing here is distinguishable from luck. The thing to improve is the evidence — more out-of-sample years per score, fewer trials spent chasing them — never the bar. Everything else in the app still works; Discover and Market are stages of a pipeline whose last gate is currently letting nobody through, and that is information.

The five surfaces

• Discover — what is worth a look? A continuous scan of the whole eToro catalog. Open it when you have no candidate. In Shariah mode it also holds ☪ Halal, the screened universe with every verdict's reasoning.
• Market — what is this thing? Chart & Levels, Analyst, Research. Open it once you have a name.
• Test — has anything ever worked here? Backtest, the microscope; Walk-Forward, the exam. Open it before believing anything.
• Portfolio — how much, and at what risk? The advised book: what cleared the bar, weighted, levered and costed. Open it weekly.
• Track — did it work? Paper books trading themselves 24/7 net of real fees, the learning verdicts, and the scoreboard for calls you recorded.

The shell, always visible: the 🔍 search blends Yahoo symbols with the ~12,700-instrument eToro catalog including the always-open 24/7 variants; the Category · Symbol pickers hold the curated universe; Interval (1m → monthly) and Range (1d → max) set what Backtest and Walk-Forward run on; the market-state chip shows open / closed / pre / after / 24-7; the selected price re-quotes about every 1.5 s and the ticker strip about every 9 s (a deliberate gap — the strip covers a whole category and hammering it gets the upstream feed to throttle us).

The Guide is not a surface. It is this drawer.

Can I trust this? — check it yourself

Don't take it on faith. The button runs a set of deterministic proofs against the live engine right now: that a position decided at a bar's close is applied to the NEXT bar, that appending future bars cannot change a past signal, that bar aggregation is causal, that costs are accounted exactly, that the Sharpe/drawdown arithmetic is textbook, that the walk-forward split leaks nothing, that a bracket fill is never optimistic, and that every risk overlay can only shrink exposure.

The gates between a pretty curve and your book — and what each one cannot do.

• The deflated bar, per score. Each score must beat its own OOS standard error (Lo 2002) times the expected maximum of N zero-edge trials (Bailey & López de Prado 2014). N counts lifetime trials, not what happens to sit in today's cache — a cache is overwritten in place and forgets everything ever tried and discarded. A score that cannot state its own OOS length is not judgeable and does not trade: the bar fails closed, because a stricter bar that silently waves through unmeasurable scores is no bar at all. It is never lowered to produce output — see where the app stands today.
• The sealed holdout. About 15% of symbols, picked by a fixed hash that cannot be re-rolled, never enter the search at all: never scored, never ranked, never promoted, never traded. Every other "out-of-sample" number here is the number selection ranks on, so it has been chosen for being large; deflation prices how many things were tried, not whether the winner generalises. Only symbols that never entered the search can answer that. A holdout result may never promote, demote or size anything — it is read by a human and by nothing else, and a test asserts the selection modules do not so much as mention it.
• Quarantine — and it is reversible. A strategy is suspended when its live expectancy is worse than its own horizon's peers and separated from them by more than noise. Never against zero: on an intraday horizon everything loses to round-trip cost, and against zero the rule convicts the whole cohort. The review runs nightly before re-conviction and releases any ban that no longer survives its own rule on clean evidence. This matters: a one-way version once banned 24 strategies that had never once reached a stop or a target.
• Evidence has to be evidence. Only a trade that ended at the strategy's own exit counts — a position closed because its book was deleted, or because a kill switch fired, pays the round-trip spread with no chance for the edge to play out. And the watcher must have been alive for the whole holding period; a stop "hit" during an outage is evidence about the outage. Because that second gate censors on the outcome — stops end fast and survive it, time exits run long and never did — the survivors are re-weighted rather than read at face value.
• The weekly sweep judges each family by its median across symbols and nine window schemes, never its best cell, and publishes what it could not judge: an absent verdict is a coverage gap, never a pass. A verdict may not be read under costs it was not computed with, so a stale one fails open — refusing to act keeps a strategy in, while acting on it would throw one out on evidence about a different system.
• The luck test (Backtest → Validate) gives a bootstrap Sharpe CI and a permutation p-value against random timing at the same exposure. It deliberately no longer answers "ROBUST": its parameters were chosen on the same series, so both statistics are anti-conservative and the verdict reads IN-SAMPLE ONLY.

What none of this promises: that a validated edge makes money tomorrow. It proves the process is honest, not that the future is known.

1 · Discover — what is worth a look?

The question: out of ~12,700 eToro instruments, which have tradeable structure right now? Open it when you have no candidate and don't want to pick one from memory.

• 🔥 At an entry spot right now — names whose geometry is at an entry, each with the plan that geometry implies: direction, entry, where to cut, and the win rate it would need to pay after fees. Re-checked every ~3 min.
• 📊 Watch board — the whole catalog ranked by pattern strength. Nothing here is at an entry yet; a name moves up to the hot list when it reaches one.
• 🎓 Validated discoveries — the exam results: which candidates the nightly walk-forward has judged, against the same bar the advised books use.
• 🛰 Coverage — every asset is either on a board or rejected with a logged reason, and the counts reconcile to the exact catalog total.
• ☪ Halal (Shariah mode only) — the full universe grouped compliant / review / excluded, each row carrying its complete justification, plus a box to screen any symbol. An automated approximation of a public standard, not a fatwa.

Distrust: a good entry spot is a fact about the chart and says nothing about whether trading it has ever paid. The gap between a long hot list here and an empty advised book in Portfolio is the funnel working, not a contradiction.

2 · Market — what is this thing?

The question: what is the shape of this instrument, and what does the rest of the system already know about it? Three views: Chart & Levels, Analyst, Research.

Chart & Levels. Everything drawn is explained by the Key strip above the candles; hide it once you know it.

• Colour is the side: teal = support (price tends to bounce up off it), rose = resistance (price tends to be capped below it) — for the diagonal trendlines and the horizontal zones alike.
• Trendlines: a solid diagonal is the line fitted through the pivots; a dashed one is its parallel channel. Log scale re-fits both in log geometry — essential for crypto and multi-year windows.
• S/R zones, labelled on the price axis: S×N / R×N, where N is how many times price has tested that level. More touches = a more respected level; ×2 is the minimum to qualify. Round numbers rank higher, because that is where real orders cluster.
• Unfilled gaps (orange dotted) act as magnets and hurdles.
• Signals overlay: stamp any strategy's trades on the candles — ▲ L long, ▼ S short, ● x exit to cash. The fastest way to see what a strategy actually does before trusting its numbers.
• Pattern callouts, MA overlays, and — on an ETORO: instrument — its real leverage tiers and fee rules above the candles.

A pivot is only usable span bars after it forms, which is why a level sometimes appears later than your eye found it. That lag is the no-look-ahead rule applied to geometry.

Analyst assembles every module's read on one name — identity, factor decomposition against the class benchmark, SEC XBRL fundamentals with the four forensic scores (Piotroski F, Altman Z, Beneish M, Sloan accruals), live technicals, dated catalysts, the real eToro cost of holding it, and the headlines. Each number ships its basis and its age, so a live quote is distinguishable from a 12-hour cache or a three-month-old filing. Two rules make it usable:

• The system's verdict comes first. Has anything this system trades ever validated out-of-sample on this symbol? "Never tested" and "nothing validated" are stated in a banner at the top, and the absence of a verdict is not a pass.
• Unavailable sections still render, with the reason. Crypto has no SEC filer; gold has no sell-side coverage. A section that vanished would read as though it did not matter.

Its entry/exit ladder shows up to four accumulation zones below spot and four scale-out zones above, each with what price must do to reach it. A zone price has already traded through is marked broken and dimmed rather than dropped — a support that failed is overhead supply now, and that it failed is information. No position sizes are emitted: tranche weights depend on account risk this module knows nothing about. Cold report ~25 s, cached 15 minutes.

The ladder draws on the real chart, not a second one. The Analyst used to build its own candle chart of the asset you were already charting — a poorer copy with no MAs, no fitted trendlines, no gaps, no signal overlay and one fixed interval. The zones and the entry/stop/target bracket now render on Chart & Levels as a layer, under the Analyst zones toggle beside S/R zones and trendlines. There is one ladder at a time and it belongs to the symbol it was computed for: change symbol and the bands come off rather than hanging over the wrong price axis, and the toggle says whose ladder it is holding. Click any row of the ladder table to jump to the chart with that band lit up.

Record a call at the foot of the report turns a read into something that can be wrong. It needs a horizon, an invalidation level, a thesis and the reference price, and it refuses without them — a view that cannot fail is not a call, and a file of unscoreable opinions looks exactly like a track record while being its opposite. When the horizon elapses the call is graded on two axes: did it pay, and did it beat simply holding the asset. The second is the one that matters. See 🎯 Analyst calls under Track for the board.

Research is the lighter view: key stats, the consensus bar and price targets, live headlines, and the instrument's eToro trading conditions. It reads the symbol from the same shared context as the chart and the Analyst, so all three Market views always describe one asset — picking a name in any of them moves the other two.

Distrust: this whole surface is descriptive. A wedge, a P/E and an analyst target are facts about the world, not evidence that trading them pays. Sell-side levels are already in the price — only revisions carry information, and high dispersion among analysts predicts low returns rather than uncertainty worth being paid for. If you record a view here, it is scored on the Track surface, and that scoreboard is what says whether your reads are worth anything.

3 · Test — has anything ever worked here?

The registry holds 2,703 strategies expanded from 98 cited base families, in nine groups: mean-reversion, trend, breakout, candles, calendar, intraday, levels, volume, ensemble. Every family carries a paper or a canonical practitioner text — hover any entry for its source. Entries like RSI(2) Pullback [deep5] are named parameter presets ("lab variants") of a family.

Backtest is the microscope. One strategy, one market, instant: equity against buy-and-hold, the metric tiles, the trade list, the entries drawn on the candles, and the luck test — a bootstrap Sharpe CI (if it spans zero the edge may not exist) and a permutation p-value against random timing at the same exposure. It runs on the top bar's symbol, interval and range. Use it to see how and why a strategy trades before spending a walk-forward on it.

Walk-Forward is the exam. Optimise on a training window, trade the next window blind, slide, repeat. Select as many strategies as you like — the run splits into batches of 100 and the leaderboard fills in between them. Click a row for the stitched OOS curve against buy-and-hold and the per-window parameter picks. If the strategy line doesn't beat the grey hold line out-of-sample, the edge isn't there.

Distrust, in the order it bites:

• A run you start here is diagnostic only. It is never persisted and can never put anything into an advised book. Nothing you do on this surface changes what trades.
• "OOS Sharpe ≥ 0.5 is respectable" is obsolete. Under real costs both classic baselines — sma_cross and donchian_break — now rate WEAK/OVERFIT themselves and are gated out of every swing pool, so "beats the baseline" is currently a very low bar. The operative gate is the deflated one in Can I trust this?, and it is far higher.
• An OOS Sharpe is not comparable across history lengths. Its null width is about 1/√(OOS years) — 0.82 at 1.5 years, 0.28 at 12.5 — so a thin score is much likelier to look good by luck. If history was short the window auto-fits and the ⤵ marker says so; a shrunk window is not full-history evidence.
• param stability floors at 1/W, one over the number of windows, so a 3-window score looks more stable than a 12-window one by construction. Read the chance-corrected figure, and read the window count beside it.
• WF efficiency is a ratio and misbehaves near zero. Under 0.5 suspect overfitting; an enormous value means the denominator collapsed, not that the strategy is extraordinary.

Every Saturday the sweep re-judges 74 base families — every pooled family the daily/swing scheme can honestly judge, derived rather than hand-picked — plus a rotating batch of 100 lab variants, across symbols × 9 window schemes, by median. Lab variants may enter an advised pool only through that promotion, and only for the horizon that earned it.

4 · Portfolio — how much, and at what risk?

Two dials and a button. Set 🛡 Risk — the maximum drawdown you accept, a hard cap that always wins over the return target — and ⚖️ smooth ⟷ returns, then press 🎯 Build my portfolio. Everything else lives behind ⚙ Advanced.

Read the result top-down: the 📋 decision card (return /mo · worst dip · straightness), then the honest walk-forward line — the same blend re-decided using only past data, and the one to trust over the headline — then the confidence grade. The trade-off menu shows the same book at five dial settings. If your filters are too strict for today's market, 🔶 best-effort mode shows the closest achievable book, labelled as such and never below the automatic bar.

• Scopes: all markets, or one book per class (Stocks, ETFs, Indices, Commodities, Crypto, Forex), plus the screened Shariah pool, the scanner's Discovered tail, and two engine scopes with fixed a-priori parameters — XS Momentum (Jegadeesh-Titman 12-1) and Pairs (Gatev et al.). Pairs currently reports itself empty because it fails its own held-out test, which is the documented post-2002 decay of that anomaly — not something to tune around.
• Horizons: Day = 15-minute bars, flat by the close · Swing = daily bars, 1–15 day holds · Long = daily bars over decades.
• Overlays only ever shrink exposure, and freed weight goes to cash, never to another position: the live regime gate 🛡, 📅 earnings blackouts (a single stock reporting within 3 days is skipped), CPI/NFP-day halving for day books, the FOMC leverage cap, and an optional correlation-cluster cap.
• Leverage follows your drawdown dial, damped by robustness and volatility, then snapped to a tier eToro actually offers — 1 / 2 / 5 / 10 / 20 / 30, there is no 3× or 7× — and never above the ESMA retail cap for that specific instrument: 30× a major FX cross and 20× a minor one, 20× a major index and 10× the rest, 20× gold and 10× other commodities, 5× stocks and ETFs, 2× crypto. Above 5× you get a warning, because margin calls are not modelled.
• Costs are real. Est. cost is the eToro round trip: spread on both sides plus overnight financing over the expected hold (weekend ×3). Scoring charges it too — per-bar carry for forex, indices, commodities and crypto, which is a CFD whenever it is leveraged or short. A red ⚠ means fees eat ≥50% of the expected move.
• The unused remainder stays in cash — cash is a position, and right now it is the whole book. Daily snapshots in reports/ keep the advice auditable.

Distrust: the headline expected return is a weighted sum of CAGRs each measured over a median ~1.5 out-of-sample years and then multiplied by leverage. On production its 95% interval ran from −16% to +50%, and from −81% to +249% at 5×. The interval ships beside the number; read it, and prefer the walk-forward line.

💾 Save & track freezes this exact combination (scope × horizon × sliders) as a book that trades itself on Track. 💾 Create N tracked books in the 🎯 panel does the same for a whole diversified set in one click.

5 · Track — did it work?

The question: does reality match what the backtest promised? Every saved book trades itself on the server on its horizon's cadence — day books every minute on 15-minute marks (forced flat within the session), swing every ~15 minutes, long hourly — filling at the latest price and logging slippage against the advised entry. Exits are stop, target, or time: day within the session (~8 h), swing 20 days, long 130 days.

Books start at $10,000 of virtual capital by default and can be denominated in USD, EUR, GBP, CHF, JPY, CAD, AUD or AED. Accounting runs in USD internally and displays at the live FX rate, so a € book moves with both its assets and EURUSD. Click a card to inspect one book — its equity curve is a first-class asset, so the whole Chart & Levels toolkit works on it.

P&L is net of real eToro fees, charged the way a real order ticket charges them: spread × exposure on each side (CFD stock/ETF 0.15%, crypto 1%), a flat $1/side commission on a real ×1 stock — free for a real ×1 ETF, and no overnight fee on either — plus overnight financing accrued per calendar night for every CFD.

✕ retires a book; it does not liquidate it. A retired book opens nothing and its unfilled limits are cancelled, but its open positions keep being managed until each one reaches its own stop, target or time exit, and the book disappears once flat. Market-closing on delete was once 78% of every fee this account has ever paid — and those round-trips taught the learning loop nothing, because a position closed administratively never got the chance to be right. Maximum fee, zero evidence. A market flatten still happens immediately if the book was ever mirrored to a broker, since a stranded broker position is worse than a fee.

The safety layer, all of it only-shrink: a per-book ⛔ drawdown throttle (entries halve past the book's dd target, pause at 1.5×), an ⚡ daily max-loss kill switch you arm per book (lose X% in a day → close everything, pause until tomorrow, push to your phone), and next to Sync now an account brake that slows every book together.

📏 The reality gap compares a book's live monthly rate against what its backtest promised at save time — but it publishes no number until the estimate can be told apart from noise. Annualising a week of returns to a monthly rate multiplies its error by about 4.3×; for a typical book the standard error only falls below the expectation it is being compared with at around three months. Until then the card shows the sample size, which is the honest answer. It used to call a book "mature" at 7 days, where the gap was roughly ten times its own noise.

🧠 Learning verdicts breaks the live evidence down per strategy — evidence count, expectancy with its confidence interval, and roughly how many more closed trades a verdict needs. Only a strategy's own exits count, and only if the watcher was alive for the whole hold. 🎯 Analyst calls is the scoreboard for views you recorded: beat buy-and-hold is its headline, not hit rate, because being long a rising market is not skill. Open calls stay visible while they run — a swing call that has not matured is the normal state, not an empty record.

Live bots: any book can mirror to a broker — simulated, eToro demo, or real — in confirm each order or auto mode. A bot can only ever mirror positions the paper book already validated, and only touches positions it opened itself. Sync now forces a pass; Reset account wipes everything back to a clean slate. Each login has its own private books, positions and history.

The daily routine (the app works even when you don't)

1. ☀️ 07:30 — the morning brief lands on your phone: account and per-book health, throttles and kill switches, the 📏 reality gaps, what the overnight learning changed, and today's macro calendar (FOMC/CPI/NFP). Nothing orange? Done for the morning.
2. During the day: pushes arrive only for new validated entries or protective events (⚡ kill switch, 🛑 brake, quarantine). No news = healthy.
3. Evening (optional): Track — the "all portfolios" candles are your combined equity; 📏 on each card says whether reality tracks the promise, once there is enough history for it to mean anything.
4. Weekly (~10 min): Portfolio → 🎯 Build → read the 📋 decision card. If weights drifted, re-save the set to rebalance.
5. The nightly report (~23:35 push) confirms the re-fit ran and what got quarantined, released or promoted.

☁️ One brain in the cloud, many screens

• The system runs 24/7 on its own server (a free Oracle machine in Paris). Your computer, your phone, your friends' browsers are all just windows onto the same live state — the same books, positions and prices everywhere.
• Nothing depends on any laptop being on. Books trade, exits fire, the 23:15 learning job runs, alerts push — all server-side.
• Exactly ONE writer, by design. Two live instances would double-trade every book. Local code changes reach the server only through a deliberate deploy; the server is the single source of truth for anything money-shaped.
• Security: zero open ports, the public address rides an encrypted tunnel, login plus invite code protect the app, and real-money trading stays triple-locked off.

The nightly loop — nobody has to remember to

• 23:15 every night: archive fresh intraday bars into the vault → re-fit every (symbol, strategy) pair in parallel → rebuild every scope × horizon book → recalibrate costs from measured paper slippage → review quarantine, then update it → promote or demote lab variants on the sweep's verdicts → push the report. Live evidence outranks backtests.
• Cost calibration may only raise friction, never lower it. The floor was once half the assumed schedule, and in production it was the outcome for all seven classes — so every walk-forward score in the system had been computed at a 2× friction discount. Until real broker fills exist, the published eToro schedule is the floor.
• Saturday: the robustness sweep re-judges 74 base families plus a rotating batch of 100 lab variants, by median across symbols and window schemes. A promotion applies only to the horizon that earned it.
• What it deliberately does not do: chase the best cell of a grid, raise leverage to hit your return target, or lower a bar to produce output.

Where the data comes from

Yahoo for stocks / ETFs / indices / FX / commodities (decades of daily bars; ~730 days of hourly, ~60 of 15-minute, ~7 of 1-minute); Binance for crypto intraday — years of 5m/15m/1h and ~30 days of 1m, ticking 24/7; eToro's own candle feed for ETORO: symbols (the prices you would actually trade, capped at 1000 bars per interval); your local CSVs. The archive vault persists every intraday bar ever fetched, so 15m history grows past Yahoo's window every night.

The data itself can leak the future, and the engine's proofs cannot see it. The no-look-ahead proofs test the code; they are structurally blind to a leak that arrives inside the feed. Yahoo's daily FX bars once carried the next session's high and low — EURUSD closed outside its own range on 2.7% of bars, USDJPY on 6.1% — and a purely bar-internal indicator correlated −0.685 with tomorrow's return. So every scored series passes one chokepoint that enforces the definition of a bar (Low ≤ min(O,C) ≤ max(O,C) ≤ High, zero fitted parameters) and refuses the feed rather than returning bad bars. Refusals show up as data excluded in a build. A dead feed — a delisted ticker serving its last bar forever — is refused too: its mark never moves, so stops and targets are unreachable.

Commodity =F symbols are continuous front-month series; roll gaps pollute long backtests, so validate on ETF proxies such as GLD.

Phone alerts

Install ntfy (free) and subscribe to the personal topic shown under 📱 Phone alerts & account settings on the Portfolio surface. Samsung: enable Instant delivery and set Battery to Unrestricted, or pushes arrive late. You get new validated entries (entry / stop / target / leverage), protective events, the ☀️ morning brief, and the nightly learning report. Day signals are scanned every minute, swing at your chosen cadence, long hourly; duplicates are never re-pushed.

Glossary

Sharpe — return per unit of risk. Drawdown — worst peak-to-valley loss; the number that makes people quit. OOS — out-of-sample: measured on data the optimizer never saw. Deflated bar — the OOS Sharpe a score must beat given its own history length and how many trials were spent finding it. WF efficiency — how much of the in-sample edge survived OOS. Expectancy — average % gain per trade. CFD — a financed derivative position; what leverage and shorts are on eToro. Book — a saved portfolio tracked in paper. Quarantine — a strategy suspended because its live expectancy is worse than its horizon's peers by more than noise; reversible, reviewed nightly. Promotion — a lab variant that earned its way into the advice through the weekly sweep, for one horizon. Sealed holdout — the ~15% of symbols nothing is ever allowed to select on. Only-shrink — an overlay that may reduce exposure and never increase it.

Phone, friends, access

One address, everywhere: https://tradingsys.tail95e19f.ts.net in any browser — no VPN, no Tailscale. Android: install /TradingSys.apk, or Chrome ⋮ → Add to Home screen. Push alerts arrive regardless of what is open, because ntfy is cloud-delivered.

Inviting someone: send them the address and the current invite code (Portfolio surface → 📱 Phone alerts & account settings, owner only). They create an account, subscribe to their own ntfy topic, build a book, and press 💾 Save & track. Point them at Where the app stands today and What this app is NOT first — in that order.

Every account is private: its own saved books, paper account, alert topic and settings. Heavy compute (the weekly sweep) stays owner-only.