CoinVantage — live crypto markets, signals & AI forecasts

Live demo: https://umer-78.github.io/coinvantage/

An installable website for live crypto markets: prices, Binance-style charts, buy/sell signals, multi-year history comparison and an AI price forecast that runs right on your machine — no paid API keys required. Works on phones and computers alike.

Signals, forecasts and AI answers are educational tools, not financial advice. The site never places trades and never asks for exchange API keys, private keys or seed phrases.

Features

Area	What it does
Markets	Top 250 coins (CoinGecko, CoinPaprika fallback), live prices via Binance WebSocket, 1h/24h/7d change, sparklines, global market cap, BTC/ETH dominance, Fear & Greed gauge, trending, top gainers/losers, heatmap, watchlist
Coin page	Live candlestick chart (1-second to weekly candles) with zoom/pan/pinch, EMA 20/50/200, Bollinger, volume, RSI, MACD, support/resistance, backtest markers and the AI forecast cone. Coins Binance does not list stream live from Gate.io, HTX or OKX over WebSocket; if the socket is blocked the chart polls every 20 s, and after a gap (a sleeping tab, a dropped socket) it fetches the missing candles instead of drawing past them
Trade signal	Buy/Sell score (−100…+100) from 9 indicator groups, multi-timeframe confluence, entry zone, stop-loss, three take-profits, exit rules and reasons
AI forecast	Leads with the 80% and 90% price ranges and how often ranges like them held in testing, then the chance of rise, shrunk by how far its direction held up on that timeframe (none on 5m, 15m, 1h, 4h and 1d, where it says so, and the ranges are centred on today's price instead of the model's lean). Where it was measured to help (5m, 1h, 4h) it also says whether a bigger or smaller move than usual is coming. An ensemble of 7 models trained in the browser on that coin's own history, each weighted by its accuracy on data it never saw
History & cycles	The long-range study: finds every past 45-day stretch whose shape matched today's, shows what price did 7/30/90 days later, overlays those charts on today's, adds year-by-year paths and monthly seasonality — and publishes how often that method was actually right on this coin
Pattern comparison	Same idea on the trading timeframe: overlays today's chart on the most similar past charts and shows what happened next
History vs now	Compares the latest 7/30/90/365 days with the previous period and the same dates 1 and 2 years ago
Backtest	Equity curve vs buy & hold, win rate, drawdown, trade list
Futures	Funding rate, annualised funding, basis, open interest and its 12h change, long/short account ratios, top-trader positioning, plus a live liquidation feed
Order book & trades	Live Binance depth and recent trades
Exchanges	Same coin on 10 exchanges (Binance, Coinbase, Kraken, OKX, Bybit, Gate.io, Bitget, HTX, Gemini, Crypto.com): price, spread, volume, best buy/sell venue
Signal scanner	Ranks the top 20/30/50 coins by signal score and the same full forecast as the coin page, on the same 1,500 candles; filters for rising, falling, oversold and overbought, and a sortable 80% range column (where price should end 8 times in 10 — the part of the forecast that tested accurate). Coins Binance does not list are read from Gate.io, HTX or OKX; coins with only CoinGecko or demo data are skipped and counted
Compare	Up to 6 coins: normalised performance, volatility, max drawdown, return/risk, correlation matrix
Tools	Coin ↔ currency converter (PKR and 15 more), a “what if I had bought regularly” DCA backtest on real daily closes vs a lump sum, and a position-size calculator from your stop-loss (risk %, fees, reward:risk). Calculators only, no trading
Trading	Three accounts side by side: demo trades you place yourself from any coin page, real trades you record (or import from a Binance trade-history CSV), and an AI running a tested strategy on simulated money with fees counted. The page states that strategy's measured result against buy-and-hold first, and the AI reviews its own trade log to switch filters on or off. A "What drove the result" panel explains any account's result in numbers: how trades ended, the win rate needed to break even, fees, the coins that cost most, and whether positions ever added up to more than the account. Nothing here places an order
AI learning	Grades every indicator reading, the forecast and the trader against what the market did next. An indicator only earns a vote after 30+ calls with a 95% range clear of 50%
Wallet	Holdings with live P&L, allocation donut, per-coin hold/exit hints; watch-only BTC, ETH, BNB Chain, Polygon and Solana addresses
Price alerts	Browser notifications while the site is open, plus server-side alerts by e-mail and Telegram while it is closed
Track record	Public page scoring every signal the server logged before the outcome was known, against the hit rate of always calling the more common outcome. A live check fits the same trust the coin page uses to the graded rows, counting coins on one candle as one observation, and says whether the live record supports showing a direction on each timeframe
News	Headlines from 5 crypto outlets, tagged by coin, refreshed every 20 minutes
AI Assistant	Chat about any coin. Answers are grounded in live data, signals, backtest, the forecast and the multi-year history study. It also explains about 40 trading terms ("what is RSI?"), how its own method works, and how to check whether a coin is a scam
Accounts	Optional sign-in keeps the watchlist, holdings, wallet addresses, chart drawings, settings and all three Trading accounts the same on every device
Premium	One-off payment for server-side alerts, VIP trade ideas and the full track record (Stripe or manual activation)
Admin panel	Traffic, users, premium grants, trade-idea publishing, site settings, API keys and manual job runs
16 currencies	Prices convert to USD, PKR, EUR, GBP, AED, SAR, INR, TRY, NGN, IDR, BRL, JPY, CNY, RUB, ZAR, BDT
Installable (PWA)	Add to home screen, works offline for the shell, app icon and shortcuts

The built-in AI (no API key)

1. Forecast engine (js/lib/predict.js) runs in a Web Worker and uses seven models:

   • chart pattern matching against the coin's history
   • similar past moves (the empirical odds after comparable moves and RSI levels)
   • gradient-boosted decision trees
   • a neural network (two-model bag)
   • logistic regression
   • nearest neighbours
   • Holt trend model

   Every model is validated walk-forward on the most recent ~36% of history. Weights come from each model's edge over a coin flip.

2. History engine (js/lib/history.js) works on daily candles so it can reach back through previous bull and bear markets. It z-normalises shape (so a $300 chart can match a $30,000 one), finds the closest past windows, measures what happened next, and then replays the whole method through the coin's own past to report its real hit rate. When nothing in history closely resembles today, it says so instead of inventing a verdict.

3. Language model (js/ai/engine.js): WebLLM runs a Qwen 3.5 model directly on the visitor's GPU via WebGPU. Free, private, cached after the first download:

   Model	Download
   0.8B	≈0.6 GB
   2B	≈1.4 GB
   4B	≈2.6 GB

   The LLM rewrites verified numbers from the analyst, which keeps small models accurate. Each model has a 4,096-token window shared by the question and the answer, so the prompt is budgeted in tokens (instructions and question, then the analyst's answer, then the last turn of the chat, then the raw data with its least useful detail dropped first) and always leaves at least 700 tokens for the reply. Every price, percentage and level in the model's answer is checked against the data it was given: one or two unknown figures are pointed out under the answer, and more than that brings back the analyst's own answer instead.

4. Rule-based analyst (js/lib/analyst.js) answers instantly on every device, including phones without WebGPU.

5. Optional: any OpenAI-compatible endpoint (for example your own Ollama server) can be set under AI Assistant → Advanced. Ollama only answers other websites when started with OLLAMA_ORIGINS set to the site's address; LM Studio needs CORS turned on.

Measured accuracy (honest numbers)

Latest (2026-10-22): the AI learns and trades in the cloud. A scheduled job grades every indicator, the forecast and the learned blend on 78 coins on every chart around the clock, into one record every device reads; another checks every signed-in account whose AI trader is on every five minutes, with the page closed, through the same check routine the page runs. Your forecast record syncs with your account, and the site keeps little on the device. The Trading page shows what each tested chart made per hour, day, week and month after real costs, beside holding the coins.

Latest: every model re-measured on all 106 coins the app can trade (2026-10-15). The coins are CoinGecko's top 150 less stablecoins, tokenised funds and coins no exchange the app reads lists (tools/universe.json names the eight left out). Each one comes from the exchange the app reads it on: Binance, Gate.io or HTX. The tests run in GitHub Actions (Evaluate workflow). A published decision changes only when two runs in a row agree. This time both runs used nearly the same candles, so the monthly re-run is the first on new data.

• Direction: no edge on any chart in either run (hit rate below the hindsight baseline everywhere). The app shows no direction call on any chart. A chart's test forecasts share their moments across coins, and crypto moves together. So 1,060 forecasts are far fewer independent tests than they look: one run 90 minutes after another moved the 1m 80% range from 69.9% to 83.3%.
• Price ranges: measured on every non-overlapping horizon, about 540,000 ranges in all, including about 162,000 on 1m. The centred 50/80/90% ranges held within 1.3 points of target on every chart (1m 50.2/79.7/89.4%). The daily range is no longer drawn wider: on 106 coins it held 81.0% as drawn.
• "Bigger or smaller move than usual": separated big moves from small ones on 5m, 15m and 1h in both runs (after a "bigger" call the move averaged 1.6–2.0× usual), so it is shown there and nowhere else.
• Score readings (1.58 million candles): on every chart the "strong rise so far" reading was followed by a rise least often (45–46%). The score describes the move so far; it is not a buy signal.
• Plan levels: with the shipped stops and targets, entering at every candle made 15m −0.004R, 1h +0.053R, 4h −0.038R and 1d −0.033R a trade. The daily plan, positive on 12 coins, loses on 106.
• Learned blend: 49.5% against the forecast's 49.6% on 5,710 forecasts, so the forecast keeps making the call.
• Outside data (Fear & Greed, funding rates, taker share, momentum, reversal): none held up.
• Score bands as strategies: each chart's best-looking score band, traded with fees against an entry at every candle, beat it only on the half of the data it was picked from and did worse on the other half (1h, 4h and 1d). No band is tradeable.
• Timing (when a move's high or low comes): 52.8% against 51.4% for a random guess over 3,822 tests. No chart's edge is bigger than the 1–4 points single charts moved between two runs, and 5m is below its baseline, so no chart's timing is labelled reliable.
• History matcher and a model trained across all coins: the matcher lost to its baseline on every chart in both runs; the cross-coin model did no better than naming the direction that was commoner in its training data (at most 1.3 points). Neither is used for a call.
• AI trader: two runs found no chart where its own rule pick cleared the bar, so every chart trades the standard rules. In the latest run holding the same coins made more than the trader on every chart; the trader fell far less (see below).

The table below is the earlier 12-coin release test, kept for comparison.

tools/evaluate-engine.mjs ran 1,440 forecasts on 12 major coins (BTC, ETH, SOL, BNB, XRP, DOGE, ADA, AVAX, LINK, LTC, TRX, DOT), 240 on each of six timeframes, re-measured on 2026-09-26. Each forecast used only data available at that moment, on real Binance candles, and is scored against the baseline of always naming whichever direction was more common in that window (a hindsight bar — nobody knows that in advance).

Timeframe	Direction accuracy	Baseline	50% range held	80% range held
1m	59.6%	63.3% (no edge)	51.7%	82.1%
5m	56.7%	58.3% (no edge)	46.3%	72.9%
15m	55.0%	60.4% (no edge)	43.8%	73.8%
1h	52.1%	55.0% (no edge)	56.7%	84.2%
4h	47.5%	59.2% (no edge)	50.8%	78.3%
1d	52.5%	54.6% (no edge)	41.3%	79.6%
All	53.9%	58.5%	48.4%	78.5%

• Direction is close to a coin flip. "Always say up" — a bar you can know in advance — scored 52.2% on the same tests. The calls where the models lean hardest were right 50.7% of the time (227 tests), no better. No model can predict crypto direction 90% of the time; any site that claims it is not measuring honestly.
• A 90% range is shown too: on the same tests the price ended inside it 87.5% of the time (1m 90.0%, 5m 82.9%, 15m 82.5%, 1h 92.5%, 4h 87.5%, 1d 89.6%).
• History as a second opinion was tested and not added: on the 240 daily tests the multi-year history engine alone scored 51.3%, a 50/50 blend with the forecast 52.5% (the forecast alone: 52.5%), and the 145 cases where both agreed 53.1%. No gain, so it stays a separate tab.
• The price range is now calibrated. It is built from each coin's own past moves (in units of the volatility at the time) instead of a normal curve × 0.75, which drew the 80% band too narrow for crypto's fat tails. Same 1,440 tests: the 80% band now holds 78.5% of outcomes (was 71.7%) and the 50% band 48.4% (was 45.1%).
• The History & cycles tab scores itself separately, per coin, and prints that number next to its verdict.
• Re-checked 2026-10-12 on fresh candles (same 1,440-forecast design): direction 1m 54.2%, 5m 52.1%, 15m 47.1%, 1h 50.8%, 4h 51.2%, 1d 43.3% — still no edge, so no direction is shown on 5m–1d. The same ranges centred on today's price (without the model's lean) held as well or better; re-measured at scale on 2026-10-13 (24 coins, about 100,000 non-overlapping ranges on 5m–1d, 5m and 15m over ~100 days) the centred 80% range held 78.5–79.8% on 5m–4h; on 1d it was narrow (75.8% of newer candles), and a width fitted on older candles (5% wider) lifted it to 77.4% on the newer ones, so the daily range is drawn 5% wider. No other chart is changed. A faster, exponentially weighted volatility made the ranges at most 1.6% better (interval score) — under the 2% bar set in advance, so it is not used; adjusting for the hour of the day (crypto is busier in European and US hours) gained at most 1.7%, also not used. A "bigger or smaller move than usual" call from volatility separated big moves from small ones on 5m, 1h and 4h (after a "bigger" call the move averaged 1.4–1.6× the coin's usual, after a "smaller" one 0.8–1.0×). The forecast's own per-coin recent check, once shown as "tested accuracy", was no guide: forecasts were right 49.5% when it read 60%+ and 49.7% otherwise, so it is no longer shown.
• Outside data was tested too (2026-10-13) and not added: Fear & Greed, futures funding rates, the taker-buy share of volume, 3-week momentum and last week's move, on 24 coins and 260 weekly dates since 2021. None held up in both halves of the data with a t-statistic of 2 (momentum worked until 2024 and then stopped), so none feeds the forecast or the AI trader.
• Treat forecasts as probabilities and always use stop-losses.

Re-run it yourself: node tools/fetch-klines.mjs klines.json then node tools/evaluate-engine.mjs klines.json 0 1 out.json.

The AI trader's rules are re-tested the same way, walk-forward: node tools/evaluate-trader-folds.mjs klines.json cuts each chart's history into four blocks and three times picks the best of 432 rule sets on everything before a block, then runs it on that block next to the standard rules, all coins in one account with fees. On all 106 coins two runs in a row found no chart where its own pick cleared the bar written in the script (on 1d it made more but fell 10 points deeper than the bar allows), so every chart now trades the standard rules (CHART_RULES is empty). On the daily chart each coin can also trade its own rules (PER_COIN_RULES, from tools/evaluate-selection.mjs, 73 coins): on the newer half of 2023–2026 they made +265.0% against +155.4% for the daily rules alone (holding: +97.3%), but a walk-forward check over three periods found them better in only one, and Settings says so. The Evaluate workflow (Actions → Evaluate → Run workflow) runs all of this on fresh candles; the latest results (2026-10-16) are on the Trading page and in CHART_TESTED and PER_COIN_RULES in js/lib/autotrader.js. In the latest run holding the same coins made more than the trader on every chart (a period's average: 15m +7.4% against +3.2%, 1h +18.2% against +8.6%, 4h +34.0% against +0.7%, 1d +44.2% against +20.5%), and a run four hours earlier on nearly the same candles had 4h at +8.3%: one run's average is rough to within several points. A published decision changes only when two runs in a row agree. The test coins are today's large coins — the ones that collapsed or were delisted are missing — which flatters every backtest here, holding's included. Fills are at the candle's price with the 0.1% fee; with fills 0.05% worse each way (spread, slippage) the trader made 0.7–2.5 points less a period. Its clearer gain is how far it fell: counting open trades at every candle close, its worst drop averaged 19.5% a period on 1d against 59.4% for holding the same coins, and 30.1% against 40.6% on 4h (on 15m and 1h it fell further than holding). The daily account beat holding in 2022–23 by sitting out the fall (−10.5% against −54.3%), not by picking winners. Each coin page shows the AI signal for its chart (buy, hold, sell or wait, from the same tested rules) with the record of buys like it on candles the rules were not chosen on: on no chart did a buy pick direction better than chance, and the page says so. Settings on the Trading page shows what the chosen chart and risk can be expected to make, read off the same walk-forward test (run at six risk levels: in the latest run more risk made more on average, always with a deeper drop and a worse worst period), and can replay any settings on recent candles. It is a demonstration, not a reliable way to beat the market.

Run locally

No build step and no dependencies: plain HTML, CSS and ES modules.

node tools/serve.mjs 8080      # then open http://localhost:8080
node --test tests/*.test.mjs   # unit tests
node tools/smoke.mjs           # loads every page in a real browser (needs playwright)

Deploy the website

It's a static site, so any static host works (GitHub Pages, Netlify, Vercel, Cloudflare Pages, S3).

GitHub Pages

1. Create a repository and push this folder to main.
2. Repository → Settings → Pages → Build and deployment → Deploy from a branch → main / / (root).
3. Wait ~1 minute; the site appears at https://<user>.github.io/<repo>/.
4. Generate the indexable pages, sitemap and robots.txt for that address and commit them: node tools/build-pages.mjs --site=https://<user>.github.io/<repo>/ (tools/deploy-push.sh runs it on every deploy)
5. Open the site → Admin → Settings → Site and paste the same URL. Alert e-mails and Telegram messages use it to link back to charts.

Everything on the free tier is fetched by the visitor's browser from public, CORS-enabled APIs, so the static site needs no server and no keys.

The backend (optional but already set up)

Accounts, cross-device sync, alerts that fire while the browser is closed, the public track record, news and the admin panel run on Supabase (Postgres + edge functions). The site degrades gracefully without it — every backend call fails soft.

• Project URL and publishable key live in js/config.js. The publishable key is safe in client code: every table is protected by row-level security, and the tables only the server may touch (admin_emails, auth_codes, rate_limits) have RLS on with no policies at all.
• API secrets (Telegram, Resend, Stripe, the cron secret) are stored in Supabase Vault and are readable only by the service role — never by a browser.
• One edge function, cv, is routed by path: /cv/account, /cv/admin, /cv/alerts-check, /cv/track-signals, /cv/news-fetch, /cv/telegram, /cv/stripe, /cv/selftest.
• pg_cron runs alerts every minute, news at :07/:27/:47, signal tracking at :05/:20/:35/:50 and a cleanup at 03:30.
• /cv/selftest returns a fixed result computed by the deployed engine. If it ever stops matching {"score":-20,"action":"SELL","probUp":0.964211}, the server engine has drifted from the browser engine.
• The whole database schema is in supabase/migrations/, in the order it was applied. To run your own copy, create a Supabase project, apply them in order (SQL editor, or supabase link then supabase db push), change the project address inside private.call_cv() (in 20260912115252_schedule_background_jobs.sql) to your own, and put your project URL and publishable key in js/config.js. Then create the two secrets the jobs and the Telegram webhook check, each a long random string: select vault.create_secret('<random>', 'cron_secret'); and select vault.create_secret('<random>', 'telegram_webhook_secret');. Without the first every scheduled job is refused; without the second the bot ignores every message.
• Sync merges each piece of a user's data separately (js/lib/cloudsync.js): whichever device changed something since the two last agreed wins. When two devices changed the same thing, the saved copy is kept, and demo or real trades only the other device had are set aside for the Trading page to offer back; two AI accounts meeting for the first time keep the one that has run longest, with trades or not (an AI account counts as soon as it starts, so a new device takes it before its first trade). Saves go through public.sync_user_data(), which only writes the keys that changed.

First-run checklist (Admin panel)

1. Sign up on the site with the e-mail that should own it, then make it an admin (SQL: update profiles set is_admin = true where email = '...';).
2. Admin → Settings → Site: public site URL and a contact address.
3. Admin → Settings → Premium plan: price, currency, period, and either enable Stripe or write the manual payment instructions shown to buyers.
4. Admin → API keys (all optional):
   • Resend API key + from-address → enables verification codes, password resets and e-mail alerts.
   • Telegram bot token from @BotFather → saving it registers the webhook automatically and the bot answers /price, /signal, /alerts.
   • Stripe secret key + webhook secret → card checkout for Premium. Point the Stripe webhook at the URL shown on that page. Checkout only returns buyers to https://umer-78.github.io (or localhost); for another domain, set the SITE_ORIGINS secret on the edge function, e.g. https://example.com,https://www.example.com.
5. Admin → Jobs: press Run now on each job once to confirm it works.

Without any of those keys the site still runs: alerts show in the browser, Premium can be granted by hand from the admin panel, and news/track record keep working.

Rebrand

Edit APP_NAME / TAGLINE in js/config.js, the <title> and meta tags in index.html, the icons (icon.svg, icon-192.png, icon-512.png, manifest.webmanifest), the link-preview image og-image.png and the site address in the og: and canonical tags and the colours in css/app.css (--accent).

Project layout

index.html              app shell
sw.js                   service worker (offline shell)
manifest.webmanifest    PWA manifest, icons, shortcuts
css/app.css             design system (dark/light), responsive layout
js/app.js               router, navigation, search, live ticker, currency/language, sync, alerts
js/config.js            app name, API endpoints, AI models, currencies, backend keys
js/i18n.js              English/Urdu strings and RTL switching
js/api/                 market data, exchanges, futures, WebSocket hub, FX rates, Supabase client
js/lib/                 indicators, signals + backtest, forecast engine, history engine, analyst, worker
js/ai/                  in-browser LLM engine and context builder
js/charts/              canvas candlestick chart, line chart, sparkline/donut/gauge/treemap
js/views/               pages: markets, coin, advice, trader, learn, scanner, track, futures, news, compare,
                        tools, exchanges, wallet, alerts, premium, account, admin, legal, auth, assistant
supabase/functions/cv/  edge function: account, admin, alerts, news, track, telegram, stripe + engine copy
supabase/migrations/    the database schema, one file per change, in the order applied
tools/                  local server, accuracy evaluation, sitemap builder, browser smoke test
tests/                  unit tests (node --test)

Data sources

• Binance public market data (data-api.binance.vision, data-stream.binance.vision) and USD-M futures (fapi, fstream)
• CoinGecko and CoinPaprika public APIs
• alternative.me Fear & Greed
• Public tickers from 10 exchanges
• mempool.space / blockstream (BTC), PublicNode RPC (EVM), Solana RPC
• open.er-api.com (currency rates, CoinGecko fallback)
• RSS from Cointelegraph, CoinDesk, Decrypt, Bitcoin Magazine, The Block

If every source is unreachable, clearly-labelled demo data is shown instead.

Known notes

• pg_net sits in the public schema because that is where Supabase installs it; the scheduled jobs call it through private.call_cv().
• get_posts is an intentionally public SECURITY DEFINER function: it is the only way anonymous visitors read trade ideas, and it nulls out the body, entry, stop-loss and targets of premium posts before returning them.

Licence

MIT licence (see LICENSE).