ETH/USD 24-Hour Trend Forecast

These labels answer one question: how likely is this upcoming economic event to deliver a "surprise" — i.e. an actual number meaningfully different from the market's consensus (forecast)?

• LOW RISK — our leading/nowcast indicator agrees with consensus, so a big data surprise is unlikely. Usually a more muted reaction.
• MODERATE RISK — the nowcast is starting to diverge from consensus; some surprise potential.
• HIGH RISK — the nowcast strongly disagrees with consensus; a notable beat/miss (and a bigger move) is more likely.

Important: this is surprise risk (nowcast vs consensus), not a guarantee of volatility. A low-surprise event can still move if the market was positioned heavily one way. It's a pre-event gauge, distinct from how far the released number ends up deviating from forecast.

Events historically able to move markets sharply — central-bank decisions (FOMC/Fed, ECB, BoE), CPI, Non-Farm Payrolls & other employment data, GDP, PMI/ISM, retail sales. We focus the War Room on these and pair them with leading indicators to flag surprise risk before they print.

Our monitor watches ETH/USD and raises an alert when price moves a notable percentage within a window (5m / 15m / 60m). "+2.61% 60m" = price rose 2.61% over the last hour. We de-duplicate these so a steady drift doesn't keep re-alerting, but a bigger accelerating leg always escalates.

Each signal is an independent measure (news sentiment, funding rates, fear & greed, on-chain flows, COT, ETF flows, Fed policy, net liquidity and more). We blend them rather than rely on any one, which reduces noise. They fall into two groups:

• Tactical (hours–days): news, fear/greed, funding, VIX, on-chain, BTC dominance, technicals, stablecoin flows, OI divergence, exchange flows.
• Strategic (weeks–months): COT, seasonal, Fed policy, ETF flows, oil, DXY, S&P 500, net liquidity.

A blended strength measure. Each indicator is weighted, decayed over time (older data counts less unless it's live), and amplified so strong signals dominate. The score tells you how aligned the market's various forces are — a high conviction score means many indicators point the same way.

Some signals (COT, net liquidity, seasonal, ETF flows) update on a slow cadence or depend on external data that isn't always available. Others are live (news, fear/greed, funding). A "No data" indicator just means there wasn't a fresh reading — we weight only what we actually have.

It tracks the ICE BofA US High Yield Option-Adjusted Spread (OAS) — the extra yield high-yield corporate bonds pay over Treasuries, i.e. a direct price on credit risk. It's the classic "risk on / risk off" canary: when credit cracks it often cracks before or together with high-beta assets like ETH.

• Headline + regime chip — the current level with a regime label: BENIGN (<3.2%), ELEVATED (3.2–4.5%), STRESSED (4.5–6.0%), CRISIS (>6.0%) — our crisis gates. For scale: the GFC reached ~21%, COVID-2020 ~11%, and 2022 ~6%.
• Signal chip + score — our blended read: + = tightening = risk-on = BULLISH for ETH, − = widening = risk-off = BEARISH. The score is a blend of momentum and position-vs-history (below).
• Δ1M/Δ3M/Δ6M/Δ1Y, 1Y Z and 1Y %ile — how fast and how far the spread is moving, and where today sits vs the last 12 months. A low %ile (e.g. 8%) means credit near its tightest of the year.
• ETH corr 90d — the 90-day correlation of daily ETH-price changes to daily spread changes. It's a descriptive gauge; it tends to turn strongly negative in stress.

Correct — the evolution is the signal. A high-yield spread is slow-moving and mean-reverting; the absolute % is largely "old news" that the market already prices. What isn't priced is the direction risk appetite is repricing right now, which is what hits high-beta ETH.

But the level isn't useless — it's the lens and the gearbox: the same +60bp widening is mild at 2.7% but tail-risk at 5.5%, because credit-to-ETH beta jumps as stress rises. So we score mostly movement (63-day and 21-day momentum, with widening weighted harder than tightening — risk-off reprices violently, risk-on slowly) and use the level relative to its own recent history (the 1Y Z/%ile) for the "situational" part. That's why today's 2.68% is read as supportive, not scary: it's near the tightest of the past year.

Practical rule of thumb: a widening spread (Δ positive, signal trending negative) is a headwind / risk-off tell for ETH; a tightening spread is a tailwind. Crossing into ELEVATED/STRESSED while momentum still points down is the canary to respect most. It's a slow, daily, regime-level input — use it as context for other signals, not as a timing trigger.

The De-bias shadow A/B panel runs three different ways of combining the same indicators, so you can see whether the blend's problem is the indicators or the maths. They all read the identical window data — only the combination rule differs.

• RAW (live) — the control, and the only arm that actually trades. It takes each indicator's level score and averages them by weight: Σ(w·s) / Σw.
• DEBIAS Z (PRIMARY) — the pre-registered candidate. Instead of the level, it asks "how unusual is this reading versus its own recent history?" Each indicator is converted to a z-score against its own trailing mean and standard deviation (z = (s − trailing mean) / trailing σ), then averaged by the same weights. It is called de-bias because it removes each indicator's permanent offset.
• DEBIAS VOTE (2°) — a deliberately cruder cousin, kept only as a cross-check. It throws away the size of the deviation and keeps just its sign (above or below its own trailing mean), then takes a weighted majority. If the z arm's edge were an artefact of magnitude weighting, the vote arm should behave differently — so it tests the hypothesis rather than assuming it.

The reason the question is asked at all: several inputs are near-constant by construction — funding rates sit in a structural contango, and fed_policy, oil_price and cot are effectively pinned. An indicator stuck at a constant contributes a permanent directional push to a raw average, not information. The de-bias arms ask whether change predicts better than level.

Baselines use prior windows only (60-window lookback, minimum 20 prior observations per indicator), so the transform never sees the window it is scoring. PRIMARY is marked in advance so the winner can't be picked after the fact; 2° means secondary. Adopting either arm is a separate, deliberate decision — the panel only gathers evidence.

The room shows each arm twice on purpose, and the disagreement is the most valuable thing on the panel.

• Full corpus (in-sample diagnostic) — every scored window ever recorded. Useful for understanding why something might work. Because the arm was chosen by looking at these same windows, it cannot be used to justify itself.
• Forward holdout (the decision set) — only windows recorded after a one-time timestamp was frozen. The transform was already fixed when those windows arrived, so this is the only figure that counts.

Worked example (as of 2026-09-28): on the full corpus the z arm looked like a large win — 53.4% vs raw's 44.0%, a paired +10.2pt with a statistically solid z≈3.75. On the forward holdout the sign reversed: −7.4pt. The in-sample win was a fitted coincidence; the honest reading is that the evidence so far favours doing nothing. If you only saw the top half of that panel, you'd have drawn the opposite conclusion.

This is why a verdict reads collecting rather than a number: the pre-registered bar is ≥30 paired directional windows, ≥+6pt, and a paired significance z≥1.3 (a statistically meaningful margin over raw). Until all three are met, the room reports progress and declines to conclude. That is intentional restraint, not a broken panel.

A win is a directional call that matched ETH/USD's 4-hour direction, scored on the snapshot recorded at the start of each 4h window. 50% is a coin flip. So a printed 43% is not "43% correct" — it is slightly worse than guessing, and that is the honest baseline this room measures against. A number worth acting on looks like a sustained margin above 50% that survives the forward holdout.

• Directional windows only — a blend too small to commit to a side reads as neutral and is excluded. Neutral calls can't be right, so counting them would unfairly penalise a cautious model. This is why the denominator (n) is smaller than the window count.
• Paired vs unpaired — a "paired" comparison only looks at windows where both arms committed to a side, so you're comparing like with like. The fixed / broke counters are that pairing: fixed = windows the alternative arm got right where raw was wrong, broke = the reverse. The net difference is the pt figure shown.
• z — a significance score, not a hit rate. z≈1.3+ means the gap is unlikely to be luck. Large pt gaps on small n are not evidence; the pt and z are read together for exactly that reason.

Because a percentage printed on 6 windows is worse than no percentage — it invites a decision that the data can't support. Cells below their minimum sample show insufficient data or collecting instead of a number. The room would rather look empty than confident, and an empty cell is not an error: it means the evidence hasn't arrived yet. If a panel has looked "collecting" for a long time, that's a signal the measurement needs a longer window — not that the market changed.

Weight is how much each indicator counts in the live blend. Suggested is a benchmark: the same weights re-normalised from each indicator's recent rolling accuracy. A positive Δ means the data would now favour a heavier weight for that indicator; a negative Δ means lighter.

Suggestions never apply themselves. An automatic 4-hourly job adapts weights, and the admin ⚡ Apply suggested weights button does it on demand — so the only honest way to read a Δ is as a suggestion, not a recommendation. Several indicators legitimately sit near zero weight (newly added inputs awaiting history, or ones that have never predicted anything), and a large Δ on a small sample is a reason for caution rather than enthusiasm.

Because the automatic weighting optimises exactly one number — recent hit rate — and an indicator whose score never moves cannot be right in a new way. It says the same thing every window, so its hit rate is simply the market's base rate for that side, not a forecast. A constant bearish vote wins 41.7% of 4-hour windows here purely because the market falls that often.

• A measured example — btc_dominance printed exactly −0.15 in 609 of 609 windows and scored 40.6%. That is the base rate to within noise, not skill. It sits inside a 55–60 band in the producer where the score is a hardcoded constant, so the underlying dominance number moving has no effect on the vote.
• It does get penalised — just gently. Measured over all 41 indicators, the three genuinely dead ones hold 7.0% of the blend at roughly 93% of the weight an informative indicator gets per unit. That is a haircut, not a retirement.
• Why it can't do better — no hit-rate threshold separates "constant" from "real but average", because a constant's hit rate is the base rate, which is an ordinary-looking number. Detecting one needs a different input (how much the score varies), not a better accuracy measure.

This is measured, not assumed — and it is why the level-shadow A/B exists: it tests the continuous version of these same inputs forward-only, against a frozen policy, before anything live changes. Nothing here adjusts the live blend.

The room shows two numbers that both used to be called "hit rate". They measure different things and they are not interchangeable:

• Indicator hit rate — pools every scored indicator row: of all the individual indicator calls ever made, how many were right. It answers "how accurate is a typical indicator call?" It is structurally pessimistic, because it is dominated by inputs whose hit rate is barely more than the market's base rate — including indicators that never vary at all (see the entry above).
• Blend hit rate — reconstructs the combined signal per 4-hour window and asks how often the thing you actually trade was right. It answers "is the blend any good?"

The blend panel also scores two baselines that require no skill at all — buy every window, and sell every window — on the same windows, so the comparison is paired. A model only earns its complexity if it beats those, and right now it does not measurably beat them:

• Live reading — blend 48.7% vs always-long 52.5% vs always-short 47.5% across 520 directional windows.
• Read that honestly. The paired test gives z ≈ −0.9, which is not a statistically meaningful gap. The truthful summary is "no demonstrated edge", not "the system is worse than random" — the confidence intervals overlap heavily.
• Why the blend is so short — it points bearish in 80.7% of windows while the market actually rose in 47.7%. That structural bias comes from near-constant bearish inputs, and it is the thing the level-shadow A/B is testing.
• Conviction does not pay. The blend's score has never once exceeded 0.31 in 646 windows, so the loud buckets are effectively empty — and accuracy is not even monotonic in magnitude inside the range that does occur.

So: use the blend number to judge the system, and the indicator number to judge a single input. This panel is measurement only — it changes no live weight or score.

What drives the tape right now names the market theme the indicators currently support — liquidity, monetary policy, risk-off, inflation, speculation, or geopolitical. A theme doesn't predict the market; it tells you which cluster of indicators the platform currently trusts, and those indicators get a temporary weight adjustment while the others are damped.

"Mixed" means no clear winner, and it is the honest answer more often than you'd expect. These themes genuinely overlap — in a typical window the top two scores are close together — so a confident label would overstate what the data supports. When there's no confident leader, the platform applies no theme tilt at all and tells you so, rather than picking one for the appearance of a view.

This panel is a recorded history, so it shows the theme as of the last recorded window — it lags the live signal by up to one refresh interval. That is by design: it's a tape of what was believed at the time, not a live re-judgement.

This is the single most misread thing in the app, so it's worth stating plainly. The sign convention is not global:

• Market Signal / indicator scores — + is bullish for ETH, − bearish.
• Economic nowcasts (CPI, labor, macro domains, ADP) — + means hotter than expected, i.e. hawkish for the Fed, i.e. bearish for ETH. A hot print shows as a positive number and a negative read for crypto. The app inverts these internally so the signal panel stays consistent, but the News nowcast cards show the raw nowcast convention.
• Credit spread — a positive score means tightening (bullish), the opposite of the intuitive "up = bad". Falling yields read the same way: lower real yields are a bullish tilt.
• War Room risk labels — LOW / MODERATE / HIGH / EXTREME describe surprise risk, not direction. HIGH RISK means an event is likely to surprise, not that it's bearish. See the War Room section above.

Rule of thumb: on the signal side, read the label with the number ("BULLISH +0.31"). On the macro-data side, read the words ("hot", "above consensus", "tightening"), because the sign there describes the economy rather than crypto.

Because old data counts for less. Each indicator has a half-life (a tactical input like news sentiment fades over hours; a strategic one like net liquidity fades over weeks), and its weight is multiplied by the fraction of that life remaining. A freshly-updated indicator contributes its full weight; a stale one contributes a fraction. So the live total usually sits a little under 100% even when every input is present, and that gap grows when a collector runs late.

Practically, this is the only reason a stale input is automatically down-weighted rather than dropped — so a "No data" indicator is not an error, just an input with less influence right now.

Several panels in the room test a candidate improvement, and they all report the same four verdicts so you can read them interchangeably:

• EDGE — the candidate beat the current model out-of-sample. The only verdict that would justify changing the live blend.
• FLAT — enough data, no real difference. Keep what you have.
• HARM — the candidate did measurably worse. Evidence against the idea.
• COLLECTING — not enough forward windows yet. No conclusion, by design.

A HARM verdict is a result, not a fault — it's the panel doing its job by ruling an idea out, which saves you shipping something that quietly degrades the signal. Note that every one of these tests is currently reporting-only: one promising idea (an agreement multiplier) was found to have no real edge and was reverted rather than left live. The room's default answer to "should we change this?" is no, not yet.

When muted or in a quiet-hours window, alerts are silent by design (the widget still updates). Also, if the app process is fully stopped (not just backgrounded), Android may delay delivery. On some devices, two alerts fired a second apart can collapse into one — the newest one always wins and the widget stays current.

We fixed this: alerts are delivered in real time via push, with a periodic backup check. Only one path posts any given alert, so you should see each alert exactly once. If you still see a double, tell us with the alert details and time — it's usually the phone re-delivering after a network drop.

No — alerts arrive as normal phone notifications. The home-screen widget is a bonus that shows the latest alert at a glance and updates automatically, so you don't have to open the app. It works alongside the notifications, not instead of them.

Open the widget app's Settings screen. You can mute alerts entirely, set a quiet-hours window (we default to 23:00–07:00) when no sound plays, and pick any of your phone's notification sounds (or Silent). The widget keeps updating even while silenced. Note: after an app update, re-allow notifications if your phone asks.

• Guest — markets, charts, news, Top Traders.
• User — trading, alarms, AI, journal, wallet, account.
• Admin — everything above plus the admin dashboard / system tools.

In the app: Account → General → the "Android Alert Widget" section has the download link and a token generator. Paste that token into the widget app's setup screen to start receiving alerts.