---
title: "Daily prognosis, 20 September 2026: three trend-led calls, and a week where our coverage is not the reason"
published: 2026-09-20T20:03:34.519019+00:00
type: prognosis_daily
canonical: https://www.moonwire.org/insights/prognosis-daily-2026-09-20.html
tags: [prognosis, forecast, calibration, regime, zcash, ethereum]
---

# Daily prognosis, 20 September 2026: three trend-led calls, and a week where our coverage is not the reason

> Bitcoin holds above its 20-day average after a 5.7% week, the same classification the standing risk-on regime read was issued under, so that call is left running rather than restated. Three new calls go out - a seven-day and a three-day lean on Zcash and a three-day lean on Ethereum - and all three are carried by price trend, with our own coverage flat or mildly against each of them. One call resolved since the last run, correctly.

## Key takeaways

- Bitcoin closed at 81,206.16 against a 20-day average of 78,473.49 after a +5.68% week - the same risk-on classification the standing regime read was issued under, so that call is left running rather than restated.
- Three new calls: Zcash higher over seven days (roughly 62% against a base rate near 57%), Zcash higher over three days (roughly 57%), and Ethereum higher over three days (roughly 55%). All three are carried by price trend; our coverage is flat or mildly against each.
- Zcash's coverage leg is three items from two analyst accounts, split one constructive and one bearish, every one a passing mention inside a piece about something else - said plainly rather than dressed up as conviction.
- One call resolved since the last run: Ethereum to beat Bitcoin over 90 days, issued 22 June at a stated 49.9%, correct (+52.6% vs +26.9%). An epoch-1 row, not pooled with current figures.
- Trailing 30 days, engine epoch 4, seven-day calls: Brier skill +0.038 on a claim pool of 14 against a 25 minimum - and a constant 'higher' forecast scored better than us over the same 14 calls. Nothing here is established.

## The window

Between 19 and 20 September the pipeline analyzed 47 items, against 8 in the preceding two-day window, when our X feed returned nothing, so the jump is a collection gap closing rather than a surge in coverage. Coin coverage in that window is led by Bitcoin at 9 mentions tagged bullish, then Solana at 3 bullish, Curve at 2 bullish, Ethereum at 2 neutral and Zcash at 2 bearish. On the sector side tokenized equities lead at 3 mentions tagged bullish, with AI at 2 tagged bearish. Five analyst accounts supply most of the volume, alongside three mover accounts.

Neither the worldwide nor the US trends list carried a single crypto topic today, so nothing entered the candidate slate from that route.

## Regime: left standing, not restated

The regime read is scored objectively, not from narrative. Today Bitcoin's daily close is 81,206.16 against a 20-day average of 78,473.49, with a seven-day return of +5.68%. Above the average and clearly outside the flat band is the same risk-on classification the standing read was issued under on 19 September, so under our append-only rule that call is left running to 26 September rather than restated with a fresh row. The older choppy read from 14 September also still runs, to 21 September, and is graded on its own window - its grade is not a comment on the risk-on call that followed it.

## Three new calls, and what is actually carrying them

The engine produces the direction and the number; we write the reasoning around it. All three of today's calls share one honest feature, so it goes at the top rather than buried: the price-trend leg is what carries each blend, and our coverage leg is flat or mildly on the other side. None of these is a multi-source conviction.

**Zcash, higher over seven days, roughly 62% against a base rate near 57%.** Zcash's coverage in the seven-day window is three items from two analyst accounts, split one constructive and one bearish with one neutral, and every one of them is a passing mention inside a piece about something else rather than coverage of Zcash itself. Centred on Zcash's own coverage history across 110 prior items, that leg carries no directional information and its residual sits slightly against the call. Note the two windows here do not agree and we are not going to smooth that over: the daily snapshot above tags Zcash coverage bearish on 2 mentions across 19-20 September, while the engine's seven-day read finds 3 items split evenly. Different windows, different counts, both stated.

**Zcash, higher over three days, roughly 57% against a base rate near 52%.** Same evidence base, scored on its own window. Under our horizon rules a three-day call on a coin already carrying a seven-day call is a separate slot, graded only against other three-day calls; it is not a restatement and not an offset. Three-day calls are a pre-registered experiment and are excluded from our published accuracy figure.

**Ethereum, higher over three days, roughly 55% against a base rate near 50%.** This one rests on more sources than the other two: 17 items in the seven-day window, 12 of them directional, from 8 distinct accounts across three source kinds, with no account past a quarter of the set. They disagree - one derivatives desk reports call-biased September options open interest while two analysts describe the same consolidation as bearish - and centred on Ethereum's own history across 800 prior items the leg nets marginally negative. Our standing seven-day Ethereum call from 19 September runs to 26 September and is graded exactly as issued.

## What we did not call, and which rule stopped each one

- **Bitcoin** convicted on both horizons, leaning higher. It is at its two-concurrent-call ceiling: a seven-day call from 15 September leaning lower and a three-day call from 19 September leaning higher, both maturing 22 September. No third row - the ceiling, not the evidence, is what stops it.
- **Ethereum over seven days and Solana over seven days** both re-convicted in the same direction as the standing rows from 19 September and inside the roughly 0.10 material-change band. Materially unchanged, so nothing is issued and both keep running.

- **UNI, CVX and XRP** each cleared the conviction gate and were then refused on evidence, all three for the same specific reason: every directional item behind the read traces to a single account, so the multi-source requirement fails. UNI has now topped or near-topped the slate for five consecutive runs on that same defect.
- **AVAX, BNB, ADA and DOGE** sit at the top of the momentum-only slate with no coverage at all behind them. They are not refusals - they simply fail the conviction gate, and we do not re-argue them each run.
- **The structural board.** All six standing question slots - the Bitcoin-record question, altseason, the majors-drawdown question, and the Ethereum, Solana and XRP relative reads - already carry two concurrent rows each, which is the per-slot ceiling. That blocks a refresh on every one of them today. It bites hardest on the Bitcoin-record card, whose two standing rows were both written by earlier engine versions and whose current read differs materially from both; we have logged that as a defect to fix rather than worked around it.
- **Events** are at their three-call ceiling.

## Receipts

One call matured and was scored since the last run.

| Call | Stated | Issued | Matured | Resolved | Outcome |
|---|---|---|---|---|---|
| Ethereum to beat Bitcoin over 90 days | 49.9% | 2026-06-22 15:09Z | 2026-09-20 15:09Z | 2026-09-20 19:51Z | **Correct** - ETH +52.6% vs BTC +26.9% |

That row was issued under engine epoch 1 and is not pooled with the epoch-4 figures below; the two are different forecasters.

One standing call deserves a note before anyone finds it themselves. Our 30 July policy-rate call is displayed with the wording "left unchanged for the remainder of 2026", and the US policy rate rose 0.25% on 16 September. The condition the row actually resolves against at year-end is narrower than that displayed wording, so the call is not yet decided - but the card and the resolution condition do not say the same thing, we do not edit published rows, and the mismatch has been logged for a fix.

## Accuracy: what the record says today

Headline, scoped: trailing 30 days, engine epoch 4, seven-day asset calls only. The Brier skill score against the base-rate reference is **+0.038** - our probabilities scored a Brier of 0.2542 against 0.2643 for simply quoting the historical base rate. That is the only figure on this page that reads the probability we actually publish, and it is marginally positive.

It is also not established. The claim pool is **14 scored calls against a 25 minimum**, so the verdict is INSUFFICIENT_DATA and nothing here supports a skill claim yet. The pool has fallen from 21 on 15 September to 16 on 19 September to 14 today, because the rolling 30-day window sheds older calls faster than the concurrent caps let us add new ones.

The Murphy decomposition says where we stand: reliability 0.026 (lower is better - this is where over-confidence would show), resolution 0.008 (higher is better - this is whether our probabilities discriminate at all) and uncertainty 0.230. A resolution of 0.008 against that uncertainty means our numbers are doing almost no discriminating work.

Two further checks fail, and we report them rather than route around them. Calibration error over the same 14 calls is 0.157 against a 10% bar: fail. And the constant-forecaster test fails - a rule that ignored every signal we have and simply said "higher" every week would have scored a Brier of 0.221 on these same 14 calls, against our 0.254. On the trailing month, that naive rule beat us. Every go/no-go bar - Brier, calibration, edge, regime and sample size - currently reads fail.

Separately, and clearly labelled: no base-rate-only answer matured in this window, so there is nothing to report in that bucket. And as a directional-accuracy diagnostic only - a statistic that is blind to the probability we publish and therefore cannot detect over-confidence, which is why it is never our headline - the trailing-30-day resolver puts asset calls at 10 correct of 17, the regime read at 2 of 4 and relative calls at 3 of 5.

The honest summary: the engine's probabilities are provisional and uncalibrated, the measured record is too thin to establish anything, and on the one month we can measure a constant guess did better. Read the reasoning, not the decimals.

---

*MoonWire surfaces and summarizes public crypto news.*

**General information only — not financial advice.** MoonWire Ltd. is not licensed by ASIC and holds no AFSL. This content does not consider your objectives, financial situation, or needs, and is not a recommendation to buy, sell, or hold any asset.

*This analysis is generated by artificial intelligence and may be inaccurate, incomplete, or fabricated. Independently verify before acting.*

© 2026 MoonWire Ltd. · [Terms](https://www.moonwire.org/legal/TOS) · [Risk Disclosure](https://www.moonwire.org/legal/RISK_DISCLOSURE) · [Privacy Policy](https://www.moonwire.org/legal/PRIVACY_POLICY)