Documentation
How it works
Lodestar runs three prediction engines side by side. Each finds trade ideas in a different way; all of them are tracked on real prices, and you decide what to act on. These pages explain every step, data source and rule, with live numbers from the engines.
V1Rules + modelA transparent rules checklist, an ML model and a news adjustment on 41 coins, every 15 minutes.
V2Shadow modelOne model tested on 8 half-years; scores the top 40 coins every 4 hours. Tracked only.
V3News engineNews first: collected, read and investigated by AI, then checked against the numbers.
The three engines at a glance
| V1 · Signals | V2 · Shadow model | V3 · News engine | |
|---|---|---|---|
| Starts from | Chart + rules score, model probability, news adjustment | Model score only (price, volume, funding) | A news story, investigated by AI |
| Checked with | Rules, vetoes, news catalysts | BTC above its daily EMA50 | Liquidity, no serious bad news, not yet priced in, reward vs risk, BTC trend; V1/V2 scores as reference |
| How often | Every 15 min, 41 coins | Every 4h, top 40 coins | Continuously, any Binance coin in the news |
| Tested on history? | Yes: no robust edge found | Yes: modest edge over 8 half-years | No (no historical news): measured live from day one |
| Alerts | None (Discord off) | None (tracked only) | Discord + tracked |
Words you don't know? Terms like ATR, EV or walk-forward are explained in the Glossary, with examples.
The question we answer
For every coin, every 15 minutes: This is the signal target. It mirrors how you trade (a target and a stop), so a prediction directly answers “would this entry have worked?”.
The pipeline at a glance
Dataprice candles, funding, Fear & Greed, derivatives, tokenomics, news
→
per coin, per 15-min candle
→
Rules score9 families, 0–100
ML modelprobability of hitting the target first
News adjustmentAI-classified events shift the probability
→
Safeguardsthresholds, vetoes, one signal per coin
→
Signalentry, target, stop, expiry
→
Track recordwin / loss / expired, automatically
Step by step
- Every 5 minutes: news. Headlines are pulled from 8 crypto news sites plus Binance listing and delisting announcements. With the news AI on, reads each new headline and turns it into a structured news event. If a strong event appears, a full scan starts immediately instead of waiting for the next slot.
- Every 15 minutes (40 s after each candle closes): market data. The latest 15-minute, 1-hour, 4-hour and daily candles for all coins and BTC are downloaded from Binance, along with funding rates, open interest, long/short ratios, Fear & Greed, CoinGecko market data and DefiLlama TVL.
- Features. numbers are computed per coin (section 5). Only information that existed when the candle closed is used; higher-timeframe values are used only once their own candle has closed. This prevents lookahead bias.
- Market-wide view. Coins are compared with each other: breadth (share of coins in an uptrend) and each coin's rank for strength vs BTC and for volume spikes.
- Rules score. Nine data families each get a 0–100 score, combined into one weighted rules score (section 6).
- Model probability. The machine-learning model turns the features into a calibrated probability (section 7).
- News adjustment. The coin's news score shifts that probability up or down (section 8).
- Decision. A signal fires only if every condition in section 9 holds. Every coin's result is stored, fired or not.
- Tracking. Each fired signal is checked against the real 15-minute candles that follow, until it hits the target (win), hits the stop (loss) or reaches the horizon (expired). That's the track record.
Data sources
| Source | What we take | Refreshed | History | Used by |
|---|---|---|---|---|
| Binance spot candles | Open, high, low, close, volume on 15m / 1h / 4h / daily; taker-buy volume | every 15 min | since 2020 ✓ backtested | model + rules |
| Binance futures | Funding rate (8-hourly) | every 15 min | since 2020 ✓ backtested | model + rules |
| Binance futures | Open interest, long/short account ratio, futures taker ratio | every 15 min | ≈30 days only (live) | rules (derivatives) |
| alternative.me | Fear & Greed index | daily | since 2018 ✓ backtested | model + rules |
| CoinGecko | Market-cap rank, market cap / FDV, BTC dominance | hourly | snapshot (live) | rules (tokenomics) |
| DefiLlama | TVL of the coin's chain or protocol and its 7-day change | hourly | snapshot (live) | rules (tokenomics) |
| News sites (8) | CoinDesk, The Block, Cointelegraph, Decrypt, The Defiant, Bitcoin Magazine, CryptoPotato, Crypto Briefing | every 5 min | live only | news score, catalysts, vetoes |
| Exchange announcements (5) | New listings and delistings: Binance, Bybit and OKX (official announcements), Upbit (Korean notices: listings, delistings, trading-caution warnings), Coinbase (new or delisted pairs spotted in its public product list) | every 5 min | live only | news score, catalysts, vetoes |
| Your LLM | with each news batch | — | news events, explanations |
“Backtested” sources have years of history, so their effect is measured on the past. Live-only sources can't be tested backwards; the track record measures them from now on.
The features the model looks at
The rules score (transparent checklist)
| Family | Weight | Tested on history | How it's scored (0–100) |
|---|---|---|---|
The rules score is the weighted average of the families that have data. A family with no data (e.g. no news in 48h) is left out, not counted as zero.
The machine-learning model
- What it is: a LightGBM model (hundreds of small decision trees). For each coin and candle it learns from the features whether the target was hit before the stop. If a candle touches both, it counts as a loss.
- What it learns from: , sampled every 30 minutes.
- Walk-forward testing: to judge it honestly, we pretend to stand at the start of each test year, train only on the past, and predict that year. Bull and bear years are both included.
- Calibration: raw scores are converted into real probabilities, so “60%” means about 6 wins out of 10 (calibration).
- Choosing the threshold: across all tested years, we measure the hit rate and expected value after fees at every probability level, allowing only one open signal per coin, as in live trading. A threshold is used only if it made money overall and in most years .
- The live model is retrained on everything except the last 6 months, which are kept aside for its calibration. Retrain it whenever you change the target or want the newest data in.
- What it relies on most: .
How news is factored in
- Classification. For each headline the AI returns: which of our coins it concerns, the event type (listing, delisting, partnership, upgrade, ETF/institutional, buyback, funding, unlock, hack, regulation…), the direction (bullish, bearish or neutral), the magnitude (1 minor, 2 notable, 3 major), its confidence (0–1) and whether it's a rumor.
- Scoring. Each event counts as
direction × magnitude × confidence × source credibility × recency. Recency halves every 12 hours. Rumors count half. The same event reported by 2+ sources within 6 hours counts ×1.3. Credibility: official exchange announcements (Binance, Coinbase, Upbit, Bybit, OKX) 1.0, CoinDesk and The Block 0.9, other major outlets 0.8, smaller outlets 0.7. The events of the last 48h add up to the coin's news score: 50 = neutral, above = bullish, below = bearish. - Probability shift. The news score moves the model's probability in log-odds:
z = (news score − 50) / 50,shift = news influence × z(now ). Example: 40% with a news score of 80 (z = 0.6) becomes . The rules score also includes news as a family (weight ). - Catalyst signals : a bullish event with magnitude ≥ 2, confidence ≥ 0.7, not a rumor, less than 6 hours old, and from an official source or confirmed by 2+ sources, can fire a signal on its own, as long as the chart score is at least 50. These are labelled “catalyst” so their results are tracked separately.
- Vetoes. A bearish hack, delisting, regulatory action, legal issue or token unlock (magnitude ≥ 2, not a rumor, last 24h) blocks any signal on that coin.
- Honest limit: free historical news doesn't exist, so the news effect can't be backtested. Both probabilities (with and without news) are stored with every signal, so the track record will show whether news improves the hit rate. Once a few months are logged, news can become a model feature too.
When exactly a signal fires
Model signal
- No veto: no severe bad news, and BTC not down more than 5% in 24h
- No signal already open on that coin
Catalyst signal
- A fresh, strong, confirmed bullish news event (section 8)
- Chart score (backtested families) ≥ 50
- No veto, no open signal on that coin
What is and isn't proven
- Proven on history: the features, the model, the rules families marked “tested” and the thresholds, all measured on years the model never trained on.
- Measured from now on: news, catalysts, derivatives (open interest, long/short) and tokenomics, through the live track record.
- Current state:
- Always: past results don't guarantee future ones. The live track record is the final judge, and signals are suggestions for you to approve, never automatic trades.
Shadow model v2 (tracked only)
A second model that came out of the September 2026 experiments. It runs next to the one above and only records its signals and their outcomes: no alerts, no trades.
- Question it answers: will the price reach +2×ATR before −2×ATR within 3 days? ATR is the coin's average daily range, so the target and stop adapt to each coin's volatility (about ±5–15%) instead of fixed percentages that sit inside normal noise.
- When and on what: every 4 hours, on the top 40 Binance coins by last month's volume (no memecoins or stablecoins).
- When it fires: the score is in the top 2% (the cut-off comes from its last validation half-year), BTC is above its daily EMA50, and no signal is already open on that coin. Signals are only created right after the 4h candle closes, so the entry is a price you could really get.
- How it was tested: 8 half-years (Sep 2022 → Sep 2026), each predicted by a model trained only on earlier data. The coins were the ones actually tradeable at the time, including 84 later delisted. It beat random picks, simple momentum, our rules score and every-candle entries. It lost money in choppy half-years and whenever BTC was below its EMA50, which is why the BTC gate exists.
- Why only tracked: the backtest edge is real but modest (about +1.4% per signal after fees, about +1% at double costs). A live track record over a few months is the honest final test.
- Retrain: about once a month, after updating the data (Data page), with the Retrain button on the V2 page.
News engine V3 (news-driven signals)
The third approach starts from the news instead of the chart: news finds the opportunity, then numbers and the models check it, and you get a report with sources and a verdict. It lives on the News (V3) page.
How a news story becomes a signal
- Collect, around the clock: exchange announcements every minute (Binance, Bybit, OKX, Upbit, plus new or removed pairs on Coinbase and OKX), 9 public Telegram channels every 2 minutes, 33 RSS feeds every 3 minutes (crypto press, SEC and CFTC, Fed and ECB, business press such as FT, CNBC, BBC), a combined Reddit feed every 10 minutes, and the Fed's meeting dates. General outlets are keyword-filtered so only crypto- or macro-related items go further.
- Triage (small, cheap AI model): every item is read once. Is it relevant? Which coins, sector or macro theme does it touch, directly or indirectly? Event type, direction, size (1–3), confidence, rumor or not, new or rehash, and any future date it mentions (which goes into the calendar).
- Stories: the same event reported by several sources becomes one story. Its score (0–100) = size × confidence × source credibility × confirmation by other sources; rumors count half, rehashes less.
- Candidates: only fresh news can become a trade, so a story is a candidate when our feeds saw it in the last 6 hours (and it isn't old news resurfacing), it is bullish and about a specific coin or sector we can trade, and it is at least "notable" (size 2, score ≥ 15). New exchange listings always qualify. Market-wide and bearish stories are context and vetoes, not candidates. Automatic investigations are paced (at most one every 90 minutes, 8 a day); a candidate waiting for a slot is shown as ITC · investigate to confirm on the watchlist and pinged on Discord, and you can investigate it immediately yourself, even past your limits.
- Deep dive (stronger AI model with web search, one candidate at a time): verifies the story at its primary source, sizes it, lists upcoming catalysts and risks, and checks whether the move already happened, using live numbers we send (price move since the story broke, move vs BTC, volume, V1 and V2 scores). It ends with a verdict: strong buy candidate, buy candidate, watch, avoid or no trade, with an expected move and horizon per coin and cited sources.
- Checks: a bullish, verified, not-yet-priced-in verdict becomes a signal only if all checks pass at that moment: tradable on Binance with at least $5M daily volume, not a memecoin, hasn't run further since the deep dive, no serious bad news in the last 24h, no signal already open on the coin, BTC above its daily EMA50 (required for "buy candidate"), and an expected move of at least 1.5 × the stop. Rejected candidates are kept with their reasons.
- Signal: entry at the live price, stop 1 × daily ATR, target = the expected move capped at 3 × ATR, expiry = the deep dive's horizon. A conviction score (0–100), a Discord alert, a full report, Explain with AI and a Markdown export.
- Tracking: each signal is resolved on real prices (1-minute candles at first, then 15-minute; a candle touching both stop and target counts as a loss).
- Learning: for every story, the price reaction of its coins is measured after 15 minutes, 1 hour, 4 hours, 1 day and 3 days (also vs BTC), plus what moved before we saw it. Over time this shows which kinds of news really move prices, and whether the AI's verdicts pay.
Why hours to days, not seconds
When a headline appears, bots and insiders often move the price within seconds. V3 doesn't try to win that race. It aims where careful reading helps: scheduled events (listings, unlocks, votes, rate decisions), news whose impact takes reasoning to see, stories that build over days, and avoiding disasters. The "already priced in?" check and the lag measurement (how much moved before our feeds saw it) keep this honest.
Cost
What is and isn't proven
- Not backtested: free historical news with reliable timestamps doesn't exist, so V3's edge can only be measured live. That's why every signal, every rejection and every story's price reaction is stored from day one.
- What to watch: the V3 track record against V2's, the "What news actually does" table (which event types move prices, and whether "buy candidate" beats "watch"), and the lag column (are we seeing news early enough?).
- Current state: Judge it after a few weeks, not a few days.
- Always: signals are suggestions for you to verify and decide on; nothing is traded automatically.