HYPE Index

Dashboard

Trade HYPE

The trading playbook — rules, real backtest cases, and the methodology lessons from the autoresearch loop. Read after How to Read.

💡

Prerequisite: this page assumes you already read How to Read. The terminology (BUY / PULLBACK / LIGHT_BUY / bright tier / Elite trifecta / late-cycle) is defined there. Here we focus on what to do with the signals, not what they mean.

Numeric claims validated as-of:2026-05-10· window: 2024-12-05 → 2026-05-09 (521 days)· script: scripts/revalidate-howtotrade-numbers.py

⚠️ Signals under observation2 pending

Recent fires where the trading rule is on probation pending a forward-return check. Live HYPE price · daily refresh.

▲ PULLBACK BUYfired 2026-05-12 · @$41.99

Concern: First-ever fire on prevState3D=RECOVERY (n=0 historical bucket). Tier finding: BEAR-prev was 100% hit historically; RECOVERY-prev is untested.

Decision date: 2026-06-02 · 12d remaining

Elapsed: 9d / 21d

Trade outcome: +47.24%

Win: ≥ +5% net at 21d → ship as-is (RECOVERY-prev validated)

Lose: ≤ -3% net at 21d → add prevState3D=RECOVERY filter

▲ LIGHT BUYfired 2026-05-14 · @$38.88

Concern: prev.bb = -1.072 — proposed `prev.bb ≥ -0.7` gate would have rejected this fire. Decision pending fwd return.

Decision date: 2026-06-04 · 14d remaining

Elapsed: 7d / 21d

Trade outcome: +59.03%

Win: ≥ 0% at 21d → keep firing, drop the gate proposal

Lose: < 0% at 21d → ship the `prev.bb ≥ -0.7` gate

Position-sizing calculator

Rule-based size with optional evidence multiplier from the signal-replay harness. Subtracts your current exposure so consecutive fires don't pyramid.

Signal typeStar qualityAccount size (USD)Current HYPE exposure (USD)Max position cap (% of account)Apply evidence multiplier (from harness backtest stats)

Recommended

30.0% of account

= $15,000 incremental

▸ Why this number? (math breakdown)

base_size[BUY][★2] = 40%

edge_multiplier = ×0.75 (n=6 < 10 (thin sample): ×0.75)

raw = base × edge = 30.0%

cap = min(raw, 40%) = 30.0%

current_exposure = 0.0% (subtracted)

final = cap − exposure = 30.0%

Floor: if final < 5%, skip the trade (signal not material enough to justify the round-trip cost).

⚠ Recommendation, not advice. Algorithm has been wrong before. Confirm thinking before acting. Sizes are incremental — if you already hold HYPE, the recommendation factors that in.

1 · Trading rules (synthesized)

Distilled from every autoresearch round. These are the rules that survived backtesting on HYPE's ~18 months of data.

📐

Rule 1 · Trade triangles, confirm with diamonds. Solid triangles are real-time and tradeable. Diamonds and late-cycle warnings have 3–6 day structural lag — they confirm what the triangle already saw.

📐

Rule 2 · Late-cycle BULL ≥8wk = size DOWN, not skip. Bull-flavor signals fired in late-cycle BULL underperformed by −36pp historically. Don't skip the trade — reduce size and tighten stops.

📐

Rule 3 · Late-cycle BEAR ≥4wk + BUY = bottom catcher. Symmetric inverse of rule 2. When the bear is exhausted (4+ weeks of weekly BEAR) and a BUY-flavor signal fires, you're catching the actual turn. 100% win on n=5 historically.

📐

Rule 4 · Elite trifecta = max size. The 3 highest- return historical longs were all Elite (1D signal + fresh 3D flip + BULL/RECOVERY family). When the purple 🔥 prints, the historical pattern is dramatic moves.

📐

Rule 5 · No SELL triangle in fast crashes — use bright −DIV. SELL signals only fire in BULL/TOPPING. If regime flips to BEAR fast, the SELL window closes before the signal can print. Bright NEG_DIV with bb-extreme covers this gap (Oct 29 2025 case).

📐

Rule 6 · Phantoms are previews, not entries. Hollow dashed diamonds in the unconfirmed zone show candidate pivots. Wait for solidification (3 bars) before treating as a signal.

📐

Rule 7 · Outline / dashed triangles are diagnostic. Suppressed BUY (outline) and low-conviction BUY (dashed) are visible for transparency — they tell you what the system filtered out. Don't trade them.

📐

Rule 8 · Stack independent timeframes for conviction. 1D + 3D agreement = strong. 1D + 3D + weekly agreement = max conviction (Elite, late-cycle bright). When timeframes disagree, the lower-timeframe is the riskier read.

2 · Backtest examples & optimization lessons

Specific cases from autoresearch where the chart's rules evolved in response to what the data showed. Each case is a real trade, the original system's behavior, what was broken, and what we changed.

Case 1 · Apr 6 2025 — the +97.8% trade we almost hid

signalBUYbb at fire−0.99fwd30+97.8%HTF🔥 Elite trifecta

What broke: the variant-C cluster filter routes only the FIRST BUY of each BEAR run into the actionable count. Apr 2 BUY had already fired, so Apr 6 was tagged suppressed — invisible despite being the highest- conviction trade in the entire validation window.

The fix:Elite override. When a would-be- suppressed BUY day ALSO has a fresh 3D BULL/RECOVERY flip, the suppression is lifted and it's promoted to a real BUY. Apr 6 is now a solid green triangle with a purple ELITE 🔥 above it.

📐

Lesson: noise filters can hide the highest- conviction trades. Always carve out an override for cases where independent confirmation is unambiguous.

Case 2 · Oct 29 2025 — the −37.6% short the SELL signal missed

signalNEG_DIV (info-only)bb at fire+0.94fwd24−37.6%weekly15wk BULL

What happened: NEG_DIV pivot fired with bb at +0.94 (euphoria). HYPE was at $48 and the pipeline expected a SELL_WARN or SELL_EXIT to follow within 30 days. Instead, regime flipped to BEAR by Nov 6 — only 8 days later. SELL signals only fire in BULL/TOPPING, so the SELL window closed before the signal could print. Price was at $30 by Nov 22.

What broke:Original chart hid Oct 29 as "info-only" because divLedToTrade === undefined — no SELL had fired in the 30d window.

The fix: bb-extreme alternate gate (NEG_DIV only). A NEG_DIV at bb ≥ +0.7 renders even without a SELL trigger, because the price extreme self-confirms the divergence. Asymmetric — see Case 5 for why.

📐

Lesson:downstream signal generators can fail in fast moves. The validation flag ("did a tradeable signal follow?") needs a backstop when the regime can change faster than the signal generator's window.

Case 3 · BUY threshold — bb ≤ −0.7 vs −0.8

Karpathy-style autoresearch sweep tested BUY thresholds across the bb_pos range. Original config used −0.7. Sweep result: −0.8 was the optimal cut on the composite metric (total_pnl × win_rate) with min_n gate.

📐

Lesson:a 0.1 threshold change can meaningfully shift signal quality. When you have data, run the sweep — don't pick on intuition.

Case 4 · The LIGHT_BUY −0.5 → −0.3 cascade disaster

The attempt: changed LIGHT_BUY threshold from bb_pos crossed up through −0.5 to −0.3 to make it more permissive (catch more soft pullbacks).

What broke: the change eliminated the Oct 2025 LIGHT_BUYs that were the only historical fires of the late-cycle BULL ⚠️ warning. The warning markers visibly disappeared from the chart even though the rule itself was unchanged.

The fix: reverted to −0.5.

📐

Lesson:cascading effects. Moving one threshold can erase visibility into a downstream feature that depends on the upstream signal's historical fires. Always check what other features feed off the data you're editing.

Case 5 · Symmetric vs asymmetric bb-extreme gate

The proposal: apply bb-extreme alternate gate symmetrically — POS_DIV at bb ≤ −0.7 AND NEG_DIV at bb ≥ +0.7 both bypass the led requirement.

The data check before applying:

Side	bb-extreme info-only	Outcome
POS_DIV	5 fires (Feb 26, Nov 22, Dec 1, Dec 9 + 1)	4 LOSERS (−31%/−17%/−18%/−8%) · 1 winner
NEG_DIV	1 fire (Oct 29)	1 winner (−27.2%)

Symmetric application would surface 4 POS_DIV losers to gain 0 winners — net regression.

The fix: asymmetric. NEG_DIV bb-extreme bypass only. POS_DIV stays gated by led-to-trigger.

📐

Lesson:never assume symmetry without verifying. The mechanics that make NEG_DIV win at bb-extreme (fast crash outpaces SELL window) don't apply on the BUY side (slow grinds give BUY plenty of time to fire). Asymmetry in mechanics implies asymmetry in filters.

Case 6 · Late-cycle BULL warning — gut feedback as hypothesis

User pain point:"sometime I long late bull and got trap."

Test: bucket bull-flavor signals by secularBullWeeksat fire. Expectation: late (≥8wk) underperforms early (<8wk).

Result (re-validated 2026-05-10):

Bucket	n	Win	Mean fwd30
LATE (≥8wk weekly BULL)	4	25%	−9.2%
Early (<8wk)	19	95%	+31.0%

Underperformance gap: −40.2pp on mean, −70pp on win rate. The original validation (n=3, −36.3pp) has held up — even strengthened slightly with the extra fire from late 2025. The pain was real, quantifiable, and reproducible on fresh data.

📐

Lesson:trader gut feedback is one of the best hypothesis sources. The validation loop's job is to formalize it into a rule, not dismiss it.

Case 7 · The symmetric mirror — late-cycle BEAR + BUY

User question:"investigate on if a strong buy after late-cycle bear, or vice versa. Is that a good signal?"

Test: bucket bull-flavor signals by secularBearWeeks at fire. Hypothesis: bear exhaustion = bottom is forming = best entries.

Result (re-validated 2026-05-10 — unchanged): n=5 fires with weekly BEAR ≥4wk returned 100% win, mean +21.2%. Same 5 dates as the original validation (Jan 16, Jan 20, Jan 24, Apr 1, Apr 2 2026) — no new fires landed in this bucket since. The mirror is real and stable.

📐

Lesson:when a warning works one direction, test the inverse. Symmetric reasoning isn't always right (Case 5) but it's always worth testing.

Case 8 · Divergences are mostly noise — Option A

User reported divergence diamonds were the noisiest signal on chart. Audit:

Tier	n	Win	Verdict
POS_DIV bright (late-cycle BEAR ≥4wk + tier-2 + led)	3	100% / +29.9%	Keep — unique signal
POS_DIV dim (led + tier-2, no late-cycle)	10	60% / +21%	Hide — redundant with triangles
POS_DIV info-only (no led, no same-day)	5	20% / +6%	Hide — genuine noise
NEG_DIV dim (led, no late-cycle)	7	29% / −4%	Hide — noise

The redundancy check: all 4 big POS_DIV dim winners (Apr 6 +104%, Aug 21 +34%, Feb 10 +29%, Feb 18 +37%) had a sibling BUY/PULLBACK/LIGHT_BUY within ±6 days. The diamond never alerted a unique trade.

Re-validated 2026-05-10 — unchanged: bright POS_DIV still n=3 (100% / +29.9%), bright NEG_DIV still n=6 (83% / −19.9%). Same fires, same outcomes — no new late-cycle divergences have landed since the original audit.

The fix: only render late-cycle bright divergences. Dim and info-only are hidden in strict mode (HYPE default). Legend still shows hidden counts for transparency.

📐

Lesson:when an overlay is redundant with a faster signal, hide it by default. Show it only when it adds unique information — the bright tier here = "not just a divergence, but a divergence at the moment of regime exhaustion when triangles might miss the turn."

Case 9 · Pivot lookforward — 7→3 day cut

The change: reduced findPivots(lookforward) from 7 to 3.

Result:

Metric	lookforward=7	lookforward=3
NEG_DIV count	9	15
SELL fires	6	8
Mean fwd30 (SELL)	−14.7%	−12.6%
Visibility lag	7 days	3 days

More fires, slightly lower per-fire quality, but 4 days less lag. Quality dilution was small; the timing improvement was material.

📐

Lesson: lag/quality tradeoffs are tunable. When the quality drop is small, prefer faster confirmation — actionable signal beats theoretical precision.

Case 10 · 5%/14d win threshold — when the test is too easy

The bug:v1 of the divergence filter scored a win as "price moved ≥5% favorably within 14d." Result: every bucket showed 100% win rate, every filter looked great.

Why:HYPE moves >5% in 14d as baseline. The threshold didn't discriminate.

The fix: v2 used r30 ≥ ±10% directional. Suddenly the buckets separated cleanly (Case 8 above).

📐

Lesson:if all your buckets win, your win threshold is too loose. A useful filter test must have failures in it — that's how you tell signal from noise.

Case 11 · Don't derive the filter from the test set

The mistake:proposed a "death-zone filter" (5 ≤ days_in_BEAR ≤ 40 AND drawdown > −45%) derived from the same backtest data it would have judged.

Caught: classic overfit. Numbers picked specifically because they happened to exclude the historical losers — guaranteed in-sample success, near-zero out-of-sample confidence.

Fix: dropped the filter.

📐

Lesson:if you derive a filter from data, you need to validate it on different data. Otherwise the backtest's "success" tells you nothing about forward performance — you've memorized the test set.

3 · Caveats & limits

⚠️

Sample size honesty.All validation numbers are on HYPE's ~18 months of post-launch data (Nov 2024 → present). Bucket sample sizes are small (n=3 to n=21). Treat percentages as directional, not as guarantees.

⚠️

Not financial advice.The chart describes historical patterns from a single token's short history. HYPE has had unique catalysts (buyback program, AF deployments) that may not reproduce. The framework is structurally sound; the specific win-rates are HYPE-specific.

⚠️

Lag is structural, not optional.Divergences need 3 forward bars; weekly streak needs week-close. You can't engineer this away — it's the price you pay for confirmation. The chart already shows you the unconfirmed zone explicitly so the lag isn't hidden.

⚠️

Composite signals can fail in regime breaks. Extreme macro events (exchange failures, hack, depeg) bypass the fee-MA classifier's assumptions. The 1D regime band is fast but not instant; expect 2–4 day adjustment lag in shock conditions.

← Back to How to Read HYPE Index →