Methodology. Every formula is open

Three methods, published side by side every day at 12:00 UTC (7:00 AM ET) and checked against the same OKX daily close. The record shows which one is closer — including when ours are wrong.

CryptoSemesh forecast

B = P + d · h   (technical base, as in v1)
S = (USD + YIELD + FED + FLOW) / 4
W = ATR14 · √h
forecast = max(0.01, B + 0.25 · W · S)   scenario: forecast ± W

Four macro factors are scored every morning from verified publications. The dollar and yield factors are computed automatically from official data (ECB euro reference rates, US Treasury daily par yield curve). Each scores +1, −1, 0 (no change or mixed signal) or n/a (no data or contradictory data, which also counts as zero):

Factor+1−1Freshness
US dollar vs euroDollar weakenedDollar strengthenedlast session, ≤ 4 days
US 10-year Treasury yieldYield fellYield roselast session, ≤ 4 days
Fed rate expectationsOdds of a hike at the next meeting fellOdds rose≤ 7 days
Spot ETF flowsConfirmed inflows or their resumptionConfirmed outflows≤ 7 days
  • If fewer than 3 of the 4 factors are confirmed, no CryptoSemesh forecast is issued for that pair that day — the record shows "Not issued" and why.
  • The factors can move the technical base by at most a quarter of W up or down. Example: base 85,000, W = 2,000, S = 0.5 → +250 → 85,250.
  • For gold (XAUT) the flow factor uses gold ETFs; for BTC, ETH and SOL — their US spot ETFs.
  • Experimental. The factor weights and the 0.25 coefficient are rules of thumb that have not been fitted on history, and there is no factor history to backtest them. Only the live record will show whether this beats plain technical analysis — that's why both are published side by side.

Technical analysis v1

b = (EMA20 now − EMA20 five days ago) / 5
d = max(−0.5 · ATR14, min(0.5 · ATR14, b))
forecast = P + d · h   scenario: forecast ± ATR14 · √h

The 20-day EMA trend projected to the close. In our backtest its error was about the same as simply assuming the price won't change (the "no-change baseline" below), and its ±ATR·√h scenario is wider than the calibrated 80% range.

CryptoSemesh range (80%)

today:  P ± 0.81 · ATR14 · √h
7 days: P ± 0.83 · ATR14 · √h
daily range ≈ ATR14 / yesterday's close
  • P — OKX spot price at 12:00 UTC (7:00 AM ET), the moment the forecast is published.
  • ATR14 — Average True Range over the last 14 completed daily candles (Wilder smoothing). What is ATR?
  • h — time to the close in days: 0.5 for today's close at 00:00 UTC, 7.5 for the 7-day horizon.
  • The coefficients 0.81 and 0.83 were fitted on 2023–2024 so that the close lands inside the range 80% of the time.
  • Model version: v2.1. Any change to the formula ships as a new version, marked in the record.

Tested on data the model never saw

PeriodBTCETHSOLXAUT
Backtest Jan 01, 2025 – Oct 04, 202679%78%80%80%
Live forecasts, 90 days80%78%80%72%

Share of days the close landed inside the range; the target is 80%. The coefficients were fitted on 2023–2024, so the backtest starts in 2025 — on data the model never saw. Every backtest day is listed in each pair's record and labeled as such. We compared ATR with GARCH, EWMA, HAR and the Parkinson estimator: they beat it by only 1–3%, so we kept the simple formula anyone can check.

Methods compared

PairCryptoSemesh forecast, avg errorTech. v1, avg errorNo-change baseline80% range hit rate
BTC/USDT 1.04% 1.08% live 1.08% 80% live
ETH/USDT 1.54% 1.50% live 1.51% 78% live
SOL/USDT 1.66% 1.61% live 1.65% 80% live
XAUT/USDT 0.63% 0.60% live 0.60% 72% live

Average error = mean |plan − actual| / actual for the daily close. "No-change baseline" is the error of simply taking the price at 12:00 UTC as the forecast — any method worth using should beat it.

Why there are no buy/sell signals

We tested a trend-following model on all of 2025: it called the next day's direction right 51% of the time for BTC and 53% for gold — no better than a coin flip. That's why the point forecasts here are published to be measured, not as trade signals, and why the calibrated range sits next to them. There are no buy/sell recommendations on this site.

Limitations

  • About 2 days out of 10 the close falls outside the range — by design. Sudden news-driven days are not predicted.
  • Pairs move together: on a bad day they can all miss at once.
  • The daily-range forecast is off by about 40% of the range itself on average, and it overestimates quiet weekends.
  • Past accuracy does not guarantee future results.