How GradedLP makes a call — and how it is judged.

A concentrated-liquidity position earns fees only while price stays inside its range. If price leaves the range (the range breaks), the position stops earning and ends up holding the asset that fell (or none of the one that rose) — impermanent loss (IL; also called divergence loss). A pool can show a high yield while positions in it lose more to price movement than they collect in fees. GradedLP measures that: for each pool or coin, over the next 14 days, how likely a position is to be pushed out of its range or to finish more than 5% behind simply holding.

We do not forecast which way the market will move. We tested it on about 15,000 historical 14-day forecasts across 19 assets: trend and momentum models were right 50.0% and 51.8% of the time — a coin flip. So no direction calls are published.

Market Weather — 14-day range forecast (34 coins)

For each coin we take its own last two years of 14-day windows and record how far price went up and how far it went down in each, keeping the two together as a pair. The chance a range breaks is the share of those windows that would have broken it. Keeping up- and down-moves together (instead of looking only at the nearer edge) made breach probabilities 18–28% more accurate in testing. Ranges shown: ±5%, ±15%, ±35%. Tiers come straight from the probability: AVOID ≥ 60%, WATCH 30–60%, OK < 30%.

The options view. For the seven coins with a Deribit options market we also record what options traders price for the next 14 days, and show it as CALM / NORMAL / STORMY. It runs as a challenger: it replaces the headline only if it beats the two-year baseline on new data for at least 120 days, by a rule sealed in advance. We say this plainly because an earlier test that appeared to show options beating the baseline was withdrawn: it had been compared against a stale baseline, and against a proper two-year baseline the edge disappeared.

SOL v2 — AVOID / WATCH / OK / OUT OF RANGE (23 Orca pools)

For each pool: how far price must move to leave the position's range, divided by that asset's normal 14-day move (a two-year figure, rebuilt on the 1st of each month on a fixed schedule and published before it is used). Less than one normal move → AVOID; one to two → WATCH; more → OK. A position whose price has already left its range by more than a day's normal move is OUT OF RANGE: it earns nothing, and usually loses nothing unless price comes back through the range. The board also shows the breach odds for a fresh position opened today.

What the history showed: on 851 past Solana outcomes, in-range AVOID calls went bad 42% of the time versus 2% for OK. The OUT OF RANGE split was introduced after seeing that out-of-range pools labelled AVOID almost never lost (6%) — an in-sample observation we disclose; only cohorts frozen after that rule was sealed are judged by it.

Poolwatch — EVM screener

A frozen scoring model selects pools on Base, Arbitrum, Ethereum and Polygon; every selection is graded 14 days later against a blind control group drawn from the same universe, so "the whole market went up" cannot pass for skill.

Pool weather boards — Optimism, XRPL, Sui, Avalanche, BNB Chain, Base, HyperEVM, Monad, Poolwatch v2

The Market Weather method, applied to each pool's own on-chain price: for a fresh ±5%, ±15% and ±35% range around today's pool price, the chance it breaks within 14 days, with the same AVOID / WATCH / OK tiers. Each chain has its own collector reading pool state every few minutes, its own daily freeze and its own sealed standard, and every call is graded on that pool's price, not an exchange price.

Where the pool's token is the real asset (matched by contract address, not by ticker — look-alike tickers exist) the forecast uses that coin's two-year history; a pool priced in the chain's native coin uses the ratio of the two coins' hourly closes; a pool with no usable history builds its own from the collector and is shown as collecting history, with no number, until it has 60 usable 14-day windows. Pegged pairs (stablecoin/stablecoin, two wrappings of one asset) are listed but not forecast.

Uniswap v4 pools (Monad, and a separate Base v4 board) are read straight from the chain's pool manager. Only pools without hooks are covered — a hook can change how the pool prices swaps or charges fees — and each pool's identity is checked by recomputing its id from its definition before it is admitted. Robinhood Chain is collecting price history only; nothing is forecast there until a standard is sealed for it.

How much evidence that is. The ten pool weather boards hold 211 pools, but they run one model, and the same coins recur: ETH is on all ten boards and BTC on eight. A calm or violent fortnight for ETH moves most boards at once. So ten verdicts are not ten independent tests — the record is read across boards, not added up, and each board's track record also shows skill against a flat rate (the best single probability per band, picked with hindsight), because a forecast can be calibrated without telling one pool from another.

Everything is checkable. Each board links to that day's pre-registration, its sealed rule and its pool list. When a window closes, the forecast is published again with its outcome and the evidence behind it: the first sample outside the band (with its block), the lowest liquidity during the window (a drained pool can't hide as "no breach"), the gap between the pool price and the exchange price, and each day's lowest and highest price with their blocks. Scripts in the public repository recompute every fingerprint and can read the pool's price on chain at the reported block. The model code is not published; each row commits to its inputs instead.

Pre-registration. Every day's calls are written to a public repository with a cryptographic fingerprint before the 14-day window opens. Anyone can check that the call existed first.

Sealed standards. The rule that decides each verdict is published and fingerprinted before any cohort it governs exists. Verdicts are mechanical: EVIDENCE OF SKILL, EVIDENCE AGAINST, UNTESTED (the cohort could not have shown the model wrong — e.g. a calm fortnight with no losses anywhere) or INVALID (data not good enough to judge). For probability forecasts the question is calibration: do 60% calls break about 60% of the time?

Confident calls count. From 2026-09-29, near-certain calls (below 2% or above 98%) are tested too, by a rule sealed in advance: if the "almost never" calls break too often, or the "almost certain" ones hold too often, that alone is evidence against the model.

Honest statistics. Pools and coins move together and daily windows overlap, so results are never treated as independent coin-tosses: uncertainty is measured by resampling whole blocks of days, and a stream needs many weeks of data before it can be called calibrated.

Solana v1 ranked risky pools correctly but under-forecast how often they trap by about 12× during a sharp August move. It has been retired and replaced by SOL v2; its published record stays.

Pool IQ (XRPL) failed its pre-registered gate and was published as FAIL (Outcome C). A new version will be pre-registered separately — not retuned on the old failure.

XRPL-MC published 16 verdicts; 14 were UNTESTED because the forecast rarely put a pool at risk. It closed on 2026-09-23; its verdicts stand, and the XRPL v2 pool weather board replaced it.

OP-MC (Optimism) published one verdict, NO EDGE. It closed on 2026-09-29; the verdict stands, and the OP v2 board replaced it.

Direction forecasting and options-based range forecasting did not beat simple baselines in testing, as described above.

The Solana record so far covers one sharp market episode; more regimes, especially a crash, are needed before any tier can claim to be proven. SOL v2 grades positions held in bands fixed on 2026-07-16, which is a held-position view, not a recommendation of a specific range. Coins with under two years of history are marked "thin history". Probabilities describe how often similar past windows broke a range; they are not guarantees. Nothing here is financial advice.

Verify every verdict and standard yourself → Public repository ↗