Activity Screener Last 24h vs each perp's own baseline
Loading…
21:00 UTC · day-ahead|23:00 UTC · final pre-open
▸ Click any row for what to expect from that symbol on the next UTC day (direction odds by hour, typical move size, row-specific adjustments). Built from a two-year study — see Screener Probabilities.
Screener Methodology & Performance
Loading backtest…
Full rule-by-rule breakdown
Rule
Picks/day
Fwd RVOL
vs market
≥1.5× hit
Fwd volat
Fwd range
Session
Retain
How to read each column — full reference
Screener
Any perp caught by at least one of four internal rules, each tuned on a 60-day backtest: Focused (≥$25M volume, RVOL ≥4× the market median, ≥3% range — ~6/day), Broad (same bar, ≥$10M — ~12/day), Emerging (≥$10M, market-adjusted volume z-score ≥3, no RVOL requirement — catches short/tight-baseline names, often newer listings, whose RVOL ratio understates their z-score), Breakout (≥$10M, RVOL ≥3× market against a baseline shifted back 3 extra days — catches multi-day moves that already pulled up their own recent baseline). ~16/day combined.
Both
Screener rows that also clear the original TradingView-style rule (day-over-day volume change ≥10%, range ≥3%, relative volume at time ≥2×). A 60-day backtest found these outperform the rest of the Screener by ~10% on next-24h volume/volatility and hit rate (95% vs 92%) — real, but a small edge, so it's shown as an extra badge rather than the default filter. Toggle Screener off and leave only Both on to see just the double-confirmed ~9/day.
Warming
Pre-surge candidates: market-adjusted RVOL between 2× and 4× (not yet a full surge) with ≥$3M turnover. In a two-year replay 8.6% became full surges the next day (11.5× the 0.74% baseline), and the hit rate climbs from 6.6% at 2–3× to 13.5% at 3–4×. It forecasts activity, not direction. Off by default.
Warming+
Warming names that also rose over the 24h, have last-4h RVOL ≥1.4× the market median, and closed in the top quarter of their 20-day range (≥0.75, when the snapshot has it). 14.4% became full surges the next day (~3 names/day in the replay). On by default.
Row dropdown
Click a row for what to expect on the next UTC day: probability the price is below the 00:00 UTC open at 1h/2h/4h/8h/12h/16h/20h/end of day, median move, typical swing, tail odds and adjustments for that row's RVOL, volume freshness, 20-day range position and stretch (24h move in its own daily σ). Probabilities come from site/data/screener_probabilities.json (see the Screener Probabilities page).
RVOL
Last 24h turnover ÷ median baseline turnover. 3× = three times a normal day for that perp.
Vol z
Robust z-score of log turnover (median / MAD), minus the market median, so a market-wide surge doesn't flag everything. Hover a cell for the raw z.
Volat ×, Volat z
The same idea on mean hourly true range %: is it moving more than usual, not just trading more?
RVOL trend
RVOL computed against the Breakout (lagged) baseline instead of the normal one. Hover for the market-adjusted z.
RVOL 4h
Last 4 hours vs the same 4 clock hours on baseline days. Below 1× means the activity has faded.
Vol DoD%, RVAT
The two numbers the original TradingView screen used — day-over-day 24h turnover change, and cumulative volume since 00:00 UTC vs the same partial-day average of the previous 10 days. Shown so you can see why a row earned the Both badge; on their own they were a much weaker signal in the backtest (2.4× forward volume vs RVOL's 8×+).
Score
0.6 × vol z + 0.4 × volat z (market-adjusted, each capped at ±6). Default sort.
Conc.
Share of the 24h turnover in the single biggest hour. Above 35% = one spike, not sustained activity.
OI Δ
Open interest change over the 24h. Rising OI + high RVOL = new positions, more likely to follow through.
Mkt cap
Circulating market cap from CoinMarketCap, the source HyroTrader says to use. Tickers are not unique, so each perp is matched to the CoinMarketCap coin whose price agrees with Bybit's (1000× contracts such as 1000PEPE use the underlying coin); a few coins CoinMarketCap lacks fall back to CoinPaprika/CoinGecko. Blank for tokenised stocks, ETFs, commodities and FX, and for coins where CoinMarketCap shows no cap. HyroTrader treats a cap below $100M as low-cap.
Avg 24h
Median 24h turnover over the baseline days — a typical day for this perp, and the denominator of RVOL. HyroTrader checks the current Bybit 24h volume (below $5M is low-cap), not this typical figure, but a typical day near $5M means the volume can drop under the line. New listings show the median of however many days exist.
low-cap
HyroTrader badge, shown when either is true: CoinMarketCap market cap below $100M, or Bybit 24h volume below $5M. Combined exposure across all low-cap names is capped at 5% of the initial account balance, including leverage. Only 40% of profit on them counts toward the profit target.
prohibited
The perp is in Bybit's Innovation Zone. HyroTrader says trading Innovation Zone assets is strictly prohibited, so these are excluded from the HyroTrader view.
stock / ETF
Tokenised stock or ETF on Bybit. The market-cap cell shows that label instead of a number, because a crypto market cap does not apply.
Executive Summary
London (12–15 UTC / 22:00–01:00 SYD) is the undisputed peak — 64% above mean volume, deepest order books, highest volatility. This is your prime trading window.
Late NY (21–23 UTC / 07:00–09:00 SYD) is the dead zone — 71–77% of mean volume, thinnest order books, highest slippage risk. Avoid market orders here.
Tokyo open (00–01 UTC / 10:00–11:00 SYD) shows the classic thin-book paradox — moderate volume but outsized price moves (110% volatility at only 100% volume).
Volume Leaders
Volatility Leaders
Full Hourly Comparison All metrics side-by-side
Visual Charts Interactive · Click legend to toggle series
Volume (Turnover USDT)
ATR(14) — Smoothed Volatility
Avg Hourly Range
Directional Bias (Signed % Change)
Descriptive average only. When hour-of-day direction is tested on separate 90-day windows it does not replicate — see the Session Edge tab.
Price Volatility |%Change|
Open Interest (Top-50 Perps)
Liquidity & Limit Order Activity Maker presence by hour
Liquidity Depth (Turnover ÷ Range)
USDT volume per unit of price movement. High = deep book = makers absorbing flow.
Amihud Illiquidity
Price impact per US$ of volume. Low = liquid (makers present). High = illiquid.
OI Change Rate — Position Churn per Hour
How fast positions open/close each hour. Higher = more rolling activity.
UTC
SYD
Session
Depth %
Amihud %
OI Chg%/hr
Interpretation
Session Analysis
Session Comparison
Session Edge Do hours and sessions predict direction? Replicated across three 90-day windows
Ranked findings
Hour-of-day direction — does it replicate?
P(basket closes up) minus that window's own all-hours baseline, by UTC hour. If an hour had a real directional edge, the three lines would sit on the same side of zero.
Each cell: P(up) (percentage points vs baseline). Combined z is the Stouffer combination across windows; ★ = passes a Bonferroni correction over all windows tested.
Asia → London
Prior NY direction → London open
Asia range → London range
Correlations with p-values. The sign of Asia→London direction flips between windows; activity links are strong in all of them.
Session open vs the previous session's range
P(up) by where the open sits in the prior range
Mean return (bp)
X axis: 0 = opens at the previous session's low, 90 = opens near its high (decile starts). Bands are 95% day-bootstrap CIs, pooled over the three windows.
P(up | opens in upper third) − P(up | opens in lower third), percentage points with 95% CI. Negative = mean reversion. Verdict "Replicates" = same sign in every window and the pooled 99% CI excludes zero.
High/low sweeps
Reversal differential = P(up | low swept first) − P(up | high swept first), measured from the sweep candle's close to session end, which cancels market drift (>0 = reversals beat continuation). "Nearer-level excess" compares the actual sweep rate of the closer level with a driftless random walk calibrated to the overall sweep rate.
Method & caveats
Rules of Thumb
Screener Probabilities What screened perps tend to do next — from a two-year study of 4,972 screened surges
Every table here is measured, not modelled: for each screened perp-day we look at the next UTC day (after the 21:00 UTC screen) and ask where price is relative to its own 00:00 UTC open after 1, 2, 4, 8, 12, 16, 20 and 24 hours. Splits use the screener's own attributes. This page is the evidence behind the "what to expect" dropdown on each screener row. It is a prior to be confirmed with order flow and price action, not a trade signal.
Bias: every split of the screened perps leans down from the 00:00 UTC open — about 56% probability below the open in the first hour, ~60% around 16:00 UTC, 58% at the close. The median name finishes 1.6% lower (2.2% after an up-surge day).
But the average is ~0: a minority pump hard (about 14% of names rise >10% by the close), so the lean is a probability tilt, not a free edge. Moves are symmetric once direction is known (median gain when up ≈ median loss when down).
It is a fade of heat, not a reversal of yesterday's direction. Down-surge days lean down too — except after a stretched drop (1–3σ), where the lean disappears.
The screener's metrics tell you how far, not which way: ATR, range, RVOL and score predict the size of the next move; none adds direction beyond the base lean.
Early warning: names at 2–4× RVOL ("Warming") become full surges the next day 8.6% of the time vs 0.74% normally; "Warming+" 14.4%.
The first 30–60 minutes carry no direction information: neither the surge day nor price/volume features predict the side of the opening-range break, and early up-breaks fade slightly more than they extend.
1 · Direction by hour
Probability that the price is below the 00:00 UTC open (P(down)), at each horizon after the open. Green–red shading is distance from 50%. ~ marks cells where the 95% confidence interval (date-clustered bootstrap) includes 50% — i.e. no demonstrable lean. Hover a cell for the interval.
Reading it: the lean is shallow in the first hours (median −0.26% after one hour — roughly cost-sized) and builds through the session, peaking around 16:00 UTC (London/NY overlap) before partly recovering into the close. The control group (non-surge ≥$10M perps) sits at 47–50% P(up) at every horizon, so the lean is specific to screened names. Small groups (n<150: fading volume, below the 20-day low, down-surge & fresh) are indicative only.
2 · How far: distribution and tail odds
Own-symbol return from the 00:00 UTC open. "σ (robust)" is 1.4826 × MAD — plain standard deviations are inflated by a handful of extreme pumps (largest: +297%, +243%, +191%), so use the robust σ and the empirical frequencies, not normal-distribution maths.
Screened names move about 2× the control's typical amount (robust σ 8.6% vs 4.3% over a day). Up-surge and down-surge days are similar in size (≈9%); flat-surge days are calmer (≈6%, 29% finish within ±2%). The distribution has a narrow core and fat tails: beyond 2 robust σ the tail frequencies are 2.4× (down) and 4.3× (up) the normal-distribution rate. The downside tail has also grown over the sample (P(fall >10% by the close) for up-surge names rose from ~12% in the first half-year block to ~24% in the latest) while the upside tail stayed ~12–17%.
3 · What the screener's other metrics add
Each screener column reconstructed at the screen and tested against the next-day return (4,972 events, two years). AUC < 0.5 means a higher value leans down; 0.5 is no information. "Size ρ" is the rank correlation with the absolute next-day move.
Direction: no metric discriminates (AUC 0.47–0.50). The "heat" metrics (RVOL, score, z-scores, last-4h RVOL) show a consistent, shallow gradient: hotter → more likely down and falls further (median −0.9% in the coolest fifth to −4.1% in the hottest by RVOL). A model using all of them has no out-of-time skill (AUC 0.51–0.52, less accurate than always guessing "down").
Size: ATR and 24h range predict how far it will move (size ρ 0.43 and 0.38; 0.53 for the live 15m ATR); an out-of-time model separates big from small movers with AUC 0.73. Use them for stop and target sizing.
Live check: on the archived screener output (20 days, 346 flagged rows) the baseline reproduces (P(up) 38.2%, median −1.76%). The 24h change is the only column with a confidence interval wholly below 0.5; funding showed no signal; the 49 rows with open interest up ≥50% were weaker (P(up) 30.6%, median −4.6% — exploratory).
4 · "Beyond its normal distribution" — stretch in σ
The surge-day move expressed in units of the symbol's own daily σ (prior 20 days), against what happens next. Tests whether a bigger stretch reverts harder.
Up-surge days
Down-surge days (|move| in σ)
Up-surge days revert at every stretch level (P(up) stays ~39–42% from 0 to 5σ) but a bigger stretch does not revert harder — except beyond ~5σ, where the median falls to −4.1%. Down-surge days are different: small drops (<1σ) keep bleeding (P(up) 40%), while drops of 1–3σ are neutral to bouncy (48–60%; the ≥2σ cells are small samples). So "mean reversion" is real for the up side and only conditional on stretch for the down side.
5 · Pre-surge: the Warming tiers
87% of surge days follow a day that was already elevated, so most surges are visible a day early. P(a full surge tomorrow) by today's market-adjusted RVOL (eligible perps ≥$3M turnover).
Warming = RVOL 2–4× the market median, ≥$3M turnover: ~10.6 names/day, 8.6% become full surges next day (11.5× baseline), stable across all four time blocks. It would have caught ~15% of all surge days (the rest come from names already ≥4×, or from <2×).
Warming+ adds: positive 24h return, last-4h RVOL ≥1.4× market, and close in the top quarter of its 20-day range → 14.4% (≈3.4 names/day). Weak signs: a negative prior day (5.3%) or fading last-4h volume (1.4%).
Coins under $10M: the tier also works (11–13% hit rate on a looser surge definition) with a smaller lift (6–8×) because their baseline surge rate is higher; their post-surge moves are about half as large.
This is an activity forecast, not a direction call: if a Warming name does surge, the section-1 profile applies (down-lean, symmetric size).
6 · The opening range does not reveal direction
Opening range = first 30 or 60 minutes after 00:00 UTC; breakout = first 15-minute close beyond it. P(price above the breakout entry at 24:00 UTC), after an up-break vs a down-break:
Nearly every day breaks (≈99%); 44% within the first hour, 76% within two hours. The surge day's direction and 10 price/volume features cannot predict which side breaks first (out-of-time AUC 0.50; 45% of screened days break up first).
The break direction doesn't predict the rest of the day either: up-breaks are followed by lower odds of finishing above the entry than down-breaks (3–4 pts), and the same gap appears on non-surge days — a general tendency to give back an opening break.
"The day closes above its open" looks very predictive after an up-break (+21 pts), but that is mostly arithmetic (price already sits above the open) and vanishes when measured from the breakout forward.
7 · Simple strategies tested (hourly bars, net of 20 bps round trip)
Range = high/low of 20:00–23:59 UTC; signal = hourly close beyond it during 00:00–07:00 UTC; exit at the opposite side of the range or 12:00 UTC.
Momentum breakouts lose in every population (worst on screened names): breaks tend to reverse. A fade of the break is the only positive result and only on screened names, but it is weak (t = 1.86), has a −110 bps median trade, one negative time block and vanishes at 40 bps costs — a lead, not a strategy. Volume confirmation and the pre-surge watchlist did not rescue the momentum version. Funding is ignored.
8 · When to screen before UTC0
The screener is published twice a day: 21:00 UTC (the canonical day-ahead screen) and 23:00 UTC (final check one hour before the open). For each screen hour, how well do the tiers anticipate a full surge on the next UTC day? (24h windows ending at that hour; 296 perps; 2024-11 → 2026-10.)
Information about tomorrow improves smoothly as the screen gets closer to UTC0 (about +0.7 points of precision per hour). Screening at 23:00 UTC beats 21:00 by about 1 point of precision and 1.4 points of recall — a small gain, so both snapshots are useful but neither is critical.
The list barely changes in the last hours: 92% of names screened at 21:00 were already screened at 18:00 UTC (99% screened-or-Warming), and 94% of those screened at 23:00 were already screened at 20:00 UTC.
The Warming hit rate (8.1–8.5%) and the direction lean (P(down) 57.7–58.2%, median −1.5% to −1.6%) do not depend on the screen hour, so the probabilities on this page apply to both snapshots.
Fresh onsets ramp from ~15:00 UTC and reach ~2× normal hourly volume by 00:00 UTC, so a name can still start surging after the 23:00 snapshot (that window was not measured separately).
9 · Reading the row dropdown
Click any row in the Screener table. The panel is generated in your browser from this page's data and the row's own columns.
Surge-day type comes from the row's 24h change: up (>+3%), down (<−3%) or flat. (The study classified the 00:00–21:00 UTC return, so this is an approximation outside the 21:00 UTC snapshot.)
Direction table: P(below the 00:00 UTC open) and the median move at each horizon for that type, shaded by distance from 50%.
How far: typical swing (robust σ), chance of falling or rising ≥5% / ≥10% by 12h and the close, and the chance of staying within ±2%.
Adjustments for that row where the study had enough data: screen RVOL tier, fresh onset vs continuation (the previous day's RVOL is inferred from the day-over-day volume change), still-hot vs fading last-4h volume, position vs the prior 20-day range, stretch in σ, open-interest change, and flags (new listing, one-hour spike).
Warming / Warming+ rows (not yet screened) show the probability they become a full surge tomorrow and which boosters they meet.
Probabilities are calibrated for the 21:00 UTC screen. The 23:00 UTC snapshot uses the same numbers: the direction lean is the same for any screen from 12:00 to 23:00 UTC (section 8).
10 · Method and caveats
Data: two years of hourly Bybit klines (2024-10 → 2026-10) for the top-300 USDT perps by turnover (296 after excluding stablecoins); 15-minute bars for the opening-range tests; archived screener output (20 days) for OI, funding and 15m ATR.
Screened = market-adjusted RVOL ≥4×, ≥$10M turnover, ≥3% range (the Broad rule). All returns are the symbol's own, from the 00:00 UTC open after a 21:00 UTC screen.
Survivorship bias: the universe is today's top perps; delisted names are missing, which probably flatters outcomes slightly.
Overlapping groups and multiple looks: ~25 splits × 8 horizons were examined and the confidence intervals (date-clustered bootstrap) are not corrected for multiple comparisons. Treat small cells as indicative.
Regime drift: downside tails grew over the sample; the blended averages understate recent downside frequency.
Not tested: order flow, BTC behaviour around the open, funding/OI as predictors over the full history (only 20 days available).
Full write-up, ideas and scripts: REPORT.md · IDEAS.md · data: site/data/screener_probabilities.json (built by research/surge_study/scripts/build_site_data.py).
Educational research on historical data. Not investment advice.
Spread & Depth Analysis Measured order-book cost per perp, and how it changes by hour — from a week of Bybit book data
The screener ranks perps by volume and range. Neither says whether the book can absorb your order: a perp can show big moves simply because a few hundred dollars of orders push a thin book around. This page measures it directly — spread, depth within a few bps of the mid, and the slippage of a market order — from a 10-second rebuild of the order book.
Median over the week. Depth is the thinner side of the book, in USD within 2, 5 and 10 basis points of the mid. Slippage is the average cost, in bps against the mid, of a market order of that size (mean of buy and sell, so it includes half the spread). Spread alone is a poor guide: a tight spread on a thin book still costs a lot to trade through.
Cost colours: under 3 bps · 3–10 bps · over 10 bps for the $50k column. A * on a depth figure means the 50 order-book levels stored by the collector do not reach that far (BTC and ETH), so true depth is larger.
2 · Thin books the archive does not cover
3 · Cost by hour of day (UTC)
Each symbol against its own average, so a value of 100 means "as usual for this symbol". The top chart is the median across all symbols. Sydney is UTC+11 in October.
4 · Best and worst hours per symbol
The three cheapest and three dearest UTC hours for each symbol, judged on its $50k market-order cost (falling back to $10k, then spread, where the book is too thin to fill that size). The ratio compares the average of the worst three with the best three.
5 · Method & caveats
Source. The perpticks collector records Bybit's 50-level order book as one snapshot per websocket connection plus incremental updates, in 5-minute files. The book is rebuilt for each symbol and sampled every 10 seconds. Spread, depth and slippage are all computed from those 50 stored levels.
Checked against the truth. Every later true snapshot was compared with the rebuilt book. Results below.
Coverage. 42 of the 67 HyroTrader-ranked perps are collected. The rest (including BTW) have no history, so section 2 shows live readings only.
One week. 12–18 Oct 2026, seven days, so one unusual day can move a symbol's hourly pattern. Weekdays and the weekend are pooled.
The 21:00 UTC hour. It was the worst hour for nearly every symbol on every one of the seven days. That coincides with the CME's daily trading pause (21:00–22:00 UTC until US daylight saving ends on 1 November); that link is an inference from the timing, not something this data proves. If it is right, the hour will move to 22:00 UTC.
Depth limit. Only 50 levels are stored, so BTC and ETH depth beyond about 1 bp is cut off.
Educational research on historical data. Not investment advice.