---
name: macro-direction
description: Study (3 Oct 2026) of who or what reliably calls the high-timeframe market direction for trading on-chain memecoins; 29 X callers scored and 30+ mechanical signals tested; the 5-state market light that came out of it. Read before any "is the market going up this week", "are we near a top/bottom", "who should I follow for macro" question.
sources: [cowork]
aliases: [macro light, market light, htf direction, who to follow macro, up wave down wave, regime light]
---

# Ask
- [stated] 3 Oct 2026: Ben wants the most reliable person or signals for high-timeframe direction (this week, next week, up wave, near top, down wave) so he can trade on-chain coins inside that direction, and find on-chain bottoms and tops.

# Data
- Targets: on-chain meme basket = equal-weight daily log returns of point-in-time CMC baskets sol_pump, sol_bonk, sol_memes, base_memes (dead coins included), Jan 2024 to Sep 2026 (idx_cmc.json from http://159.223.88.168/hx-c9ois/edge/chain-baskets-2026-10-01/all.tgz); longer check = OKX listed-meme index (DOGE, SHIB, PEPE, FLOKI, BONK, WIF, BOME, MEW, PNUT, GOAT, PENGU, TRUMP, MOODENG, NEIRO, TURBO, ACT; first 30 days after listing dropped), Sep 2020 to Oct 2026 (survivor-biased); BTC Coinbase daily 2017 to 2026.
- Signals: BTC MAs and momentum, meme own trend, ETH, SOL/BTC, Fear and Greed (alternative.me), BitMEX funding (full history), stablecoin supply (DefiLlama), pump.fun fees and Solana DEX volume (DefiLlama), SPX, NDX, DXY, 10y (Yahoo).

# People (29 X accounts, Jan 2024 to Oct 2026, 3,493 directional calls, top 3 market tweets per week each)
- Scored vs same-direction calls on random dates within 60 days. All callers together: 28d hit 50.6% vs 51.6% null; strong-conviction calls (n=362) worse than random (-1.9 pts). Crowd bullishness across them: corr +0.1 with next week, not significant, mostly follows trend.
- Small positive, same sign both halves: cburniske (n=58, +10 pts at 28d, HTF/months, infrequent), Ninjascalp (n=68, +19 pts, contrarian and sarcastic, hard to read live), Tradermayne (+6), CryptoCapo_ (+7), CryptoHayes (+6), PeterLBrandt (+6), _Checkmatey_ (+6, only late half). None proven: with 29 tested, one or two 2-sigma results are expected by luck.
- Zero or negative: AltcoinSherpa, MacroCRG, EmperorBTC, IncomeSharks, ki_young_ju, ColdBloodShill, inmortalcrypto, rektcapital, MustStopMurad, Bluntz_Capital, CredibleCrypto, blknoiz06 (Ansem), Rewkang, CryptoKaleo, krugermacro, DaanCrypto, caprioleio, CryptoTony__; Pentosh1 too few calls (14); HsakaTrades 3. KingCobra scored separately: random ([[kingcobra-trades]]).
- In 2024 to 2026 BTC was mean-reverting at the 28d horizon: a dumb "bet against the last 4 weeks" rule beat every caller's contrarian calls (+15 pts vs +9).

# Mechanical findings (on-chain basket 2024 to 2026 unless noted)
- BTC above its 50-day average = the backdrop that matters. All 48 threshold variants agree: memes do better next week when BTC is above, on both datasets.
- Memes rallying while BTC is below its 50-day average (counter-rally) = trap: next week -4.5% avg on-chain, -7.2% on the 2020 to 2026 index, worst state.
- Froth flags (any one: pump.fun fees 7d avg > 1.5x 28d avg; SOL/BTC +20% in 4 weeks; meme index doubled in 4 weeks) = next month -13% avg, 44% chance of a -30% drawdown vs 26% base; 48 of 48 variants worse on-chain (32 of 48 on the listed index).
- Washed out (meme index -40% in 4 weeks) is NOT a bottom: 54 to 58% chance of another -30% within a month.
- Fear and Greed < 20 or negative funding: on-chain basket rarely fell another 30% next month (1 to 3% vs 26%) but upside was also flat; did not replicate on the 2020 to 2026 index. TESTING only.
- BTC trend filters do not beat holding BTC in bull or chop years (2017 to 2020, 2024 to 2026); they cut the big bear drawdowns (2021 to 2023 x2.5 to x3.9 vs x1.4). Insurance, not a return booster.
- Weekly direction of memes is mostly noise: medians are negative in every state (baskets bleed); the light shifts the mean and the crash odds, it does not make weekly calls reliable.

# The 5-state light (weekly, Sunday close)
- 1 UP-WAVE: BTC > 50d avg and memes > 20d avg. Next week +3.9% avg on-chain, next month +16% avg.
- 2 DIP IN UPTREND: BTC > 50d avg, memes < 20d avg. Best entry zone; next month +15% avg (2020 to 2026 index +32%).
- 3 COUNTER-RALLY: BTC < 50d avg, memes > 20d avg. Fade it.
- 4 DOWN-WAVE: BTC < 50d avg, memes < 20d avg. Stay small.
- 0 HOT: any froth flag. Take profits; overrides the others.
- 3 Oct 2026 reading: UP-WAVE (BTC 84.6k vs 50d avg 78.6k, 4w +6%; memes +3% over 20d avg; pump.fun fees 1.32x, SOL/BTC +9% 4w, F&G 67). Score next week.

# Raw data (this Cowork session scratchpad, temporary): macro/daily2.pkl, signals2.pkl, people/*.json, pb/*.out.json, people_calls.pkl, scripts sig2.py, light2.py, robust.py, pscore.py, base.py.

# On-chain risk-on/off + rotation system v1, 3 Oct 2026
- [stated] 3 Oct 2026 Ben: he is very bullish, buys, makes money, then a down wave he does not see takes weeks and all the profit. Wants a system that fixes that: market-wide risk on/off (BTC extended about to pull back hurts on-chain; BTC turning up = risk on start) plus rotation between hot baskets (e.g. LONG eco topping) and under-the-radar baskets turning (e.g. on Solana). Priority Robinhood Chain, then Solana. Believes sentiment (timeline super bullish = top; "scam, dead, PVE" = bottom) plus volume times tops and bottoms.
- [stated] Ben asked whether the BTC 50d average is always support in a bull market.
- BTC 50d answer: 2017 to 2026, BTC sat below its 50d avg on 27% of bull days (above 200d); 47 bull-time breaks, half reclaimed within 7 days, half fell another 10%+. Not support, not a timing signal.
- BTC pullbacks are NOT predictable from stretch: P(BTC -5% within 10d) any day 38%, >8% above 20d avg 34%, RSI>70 28%, RSI>80 25%. Stretch even favours baskets (more +25% first than -25%).
- But the hit is real and same-week: BTC week < -5% -> Solana baskets -12% to -17% that week (n 12 to 20); BTC week > +5% -> +11% to +25%. Beta ~3x. After a BTC -5% in 3d, basket down-wave odds are LOWER (8% vs 15%): the hit is already in.
- Robinhood: RH memes basket correlation to BTC only +0.23 (Aug to Oct 2026); RH down wave was chain-internal: Pons V2 fees 7d avg went from $0.2M (mid Aug) to $8.1M (10 Sep peak), RH memes peaked 3 to 10 Sep, fees then fell 83% and RH memes fell 64%. X mention rate for RH peaked 7 Sep (160 posts/h sample) at the top.
- Launchpad fee state test (Solana, 2024 to 2026): fees 7d avg 40%+ under 45d peak (COOLING) -> pump basket next 21d -13% avg, -25% first 34% vs 13% +25% first; fee MANIA (7d > 1.5x 28d) -> own basket -12% to -15% next 21d (pump, bonk).
- SOL/BTC +20% in 4 weeks -> Solana basket -25% first 28% vs +25% first 4% (7:1). F&G > 75 doubles down-wave odds but upside equal. Basket -40% from 30d high = knife (down 30% vs up 20%).
- X attention test (twitterapi.io daily samples 2024 to 2026, 2 x 20 latest posts/day; rate = posts per hour from timestamps; mood = bull-word share minus bear-word share): mention spike (7d z > 1.5) -> pump basket -13% to -19% next 21d, bonk -24% (both halves negative); spike + bullish mood -> -25% first 52 to 63%; quiet (z < -1) -> down-wave 14 to 17% vs 25 to 28%; quiet + bearish -> down-wave 1 to 10%, +25% first 26 to 34% (the floor). Loud + bearish (scam chatter while busy) is NOT a bottom. 73 to 76% of 72 parameter variants agree; strongest on pump and bonk.
- EXIT test (biggest fix): basket 30d breakout entries, 2024 to 2026: hold 5 days PF 2.20 (9 baskets) / 2.95 (Solana) / 3.59 (2020 to 2026 index) vs v3 exit at 15d low PF 1.13 / 1.80 / 2.03; trend-break exits give back ~19% from peak; trailing -30% worst (PF 0.60). Breakouts are 1 to 2 week bursts; take profit within 5 to 8 days. Entry filters (skip chain COOLING, BTC>20d) did not add on top (small n).
- Rotation test: equal-weight point-in-time baskets bleed (x0.18 since May 2024); only "hold baskets already hot next week" made money (x1.52); turning-from-base rules lost early, won late. Baskets = weather gauge for picking coins, not the instrument.
- Live read 3 Oct 2026 ~12:50 Bangkok: BTC 84.6k above 20d (82.4k) and 100d (71.0k) = market risk on. Solana: X memes mention z +3.6, mood z +3.1 (rate 180/h mid Sep -> 530 to 840/h since 26 Sep) = LOUD + BULLISH top warning; pump.fun fees 1.32x 28d (not mania yet); letsbonk COOLING (-41% vs peak); SOL/BTC +9% 4w. Robinhood: Pons fees COOLING (-83% vs 10 Sep peak); RH X mentions z +1.6, mood z -1.1 = loud and bearish, not the quiet floor yet. Score 24 Oct.
- Raw (this session scratchpad, temporary): macro/xs_memes.json, xs_sol.json, xs_rh.json, sentx.py, senttest.py, fees.py, exits*.py, system.py, rot.py, lead.py, t1.py, fee_*.json, pons.json.

# v2 re-check and new layers, 3 Oct 2026 afternoon
- [stated] Ben: anger (loud and angry) is not the bottom; depression (gone quiet, called dead) is. Wants LunarCrush checked (influencers, sentiment, coin engagement), the breakout clock defined and tested, Trader XO's "alt phases last 7 to 8 weeks" tested, the big-then-mid-then-small coin rotation tested, bottoms tested as 50 to 70% pullbacks plus untested HTF order blocks plus RSI divergence plus 100/200 MAs plus sentiment, "on-chain sells off first" order tested, money flows, and narrative leaders (e.g. @theunipcs, @insomniacxbt, @Natan_benish) read as a separate layer from emotional crowd sentiment. Dashboard/Telegram only after the system is tuned.
- CORRECTION: the 3 Oct "Solana loud + bullish" reading was a data artifact. twitterapi.io search density doubled for every topic from about 26 Sep (control query coffee/breakfast 7d rate 290 to 330/h then 723/h on 2 Oct). Rates are now divided by that control. After the fix: Solana memes mention z +1.3 (not loud), RH -0.5.
- Re-test with control-normalised X rates (72 variants): loud + bearish (anger) worst, 90% of variants negative, 72% both halves, -13% excess next 21d; loud alone 68% negative; loud + bullish 65% negative (-8%); quiet alone 88% positive (+4%); quiet + bearish weaker than first reported (+3%, 43% negative). Floor = quiet, not angry.
- LunarCrush MCP is in limited mode: 30 days of daily data only, no backtest possible. Anecdote: RH engagement 1-year high 3 Sep 2026 = RH memes top; RH mentions high 12 Jul (launch). LunarCrush sentiment % sits 75 to 94 always, not useful. long.xyz topic posts 229 (13 Sep) to 19 to 36/day (1 to 3 Oct), sentiment 37 on 1 Oct.
- Breakout clock: day 0 = first daily close above the prior 30-day highest close, with BTC above its 100d avg. Median trade vs random day peaks around day 8 to 13 (Solana +5 to +7% vs -2 to -6%), median turns negative by day 21 to 34; mean keeps rising to day 34 from a few huge runners. Sell schedules: all out day 13 (Solana median +7.4%, PF 2.36) or half day 8 + rest by 15d low max 55d (avg +11%, PF 2.58); all out day 55 median -15%. Base AI/memes breakouts negative even at day 8.
- Trader XO 7 to 8 weeks: up-phases (first close above 50d avg after 20 of 30 days below; end at 25 to 40% drop) are bimodal: failed ones peak in 0 to 25 days, big ones in 100 to 200 days. Median start to peak: on-chain Solana 2 weeks (n=8), listed memes 4.3 weeks (n=14), 30 OKX alts 3.6 weeks (n=17). Holding 8 weeks from start: on-chain median -13%, alts +1%, listed memes +4%. Fixed 7 to 8 weeks does not fit on-chain.
- Size waves (Ben's big then mid then small): RH (91 coins, mcap tiers top 5 / 6 to 20 / rest) BIG and MID peaked 3 Sep 2026, SMALL 18 Sep, then all fell. Solana (81 coins, Dec 2025 to Oct 2026, top 10 / 11 to 40 / rest): BIG rolled below 20d avg after a 30d high -> chain index next 21d -13% (n=17, both halves negative); BIG rolled over while SMALL still rising -> -18% (n=9, both halves); SMALL beats BIG by 20%+ over 14d -> -7%; all three tiers above 20d avg -> +6% (n=57). TESTING, short history.
- "On-chain sells off first": weekly lead-lag correlations about 0 (2024 to 2026); at alt-index tops (25% swings) on-chain memes topped a median 10 days earlier. Weak, not a timing tool.
- Bottom confluence (44 coins: 95 RH via GeckoTerminal, Solana pump/bonk, 49 OKX listed, bars 50 to 80% under ATH after a 3x run, one per coin-day): depth 70 to 80% better than 50 to 70% (+50% first 50% vs 32%); RSI 4h<30, untested 4h/12h OB, RSI divergence each cut the odds of another -25% leg (10 to 28% vs 38%) but did not raise upside; untested HTF OBs inside the zone are rare (10 cases); count of confluences 0/1/2/3+ not consistent across halves or chains. Ben full set (fib + OB + RSI) n=12, no edge shown. Confirms [[reversal-zones]]: depth and RSI matter, exact lines do not.
- Narrative leaders (theunipcs 210, insomniacxbt 116, Natan_benish 57 classified posts, 2024 to 2026): EUPHORIA posts -> topic index next 21d median -13% (base -7%); REASSURE after a 15% drop -> median -9% (no bottom); WARN -3%; leader CAPITULATE n=2 -> +47%. unipcs 2 Oct 2026 RH thread = REASSURE (cooldown normal, next leg soon, Robinhood listings the fix).
- Live 3 Oct 2026 afternoon Bangkok: AI 0x2e8c31162b855a2ffa90f6f8634643ad6f111e18 (pool 0xc4a21f9d6485fc5893dd4a491b320a83daf4da1d) top $381M 17 Sep, now $128M (66% off, 0.66 fib $130M, 0.786 $83M), RSI 4h 34 / 12h 31. PONS 0x39dbed3a2bd333467115de45665cc57f813c4571 (pool 0x10cc6bd38112cac182db90b6a71d8bb5939526ba) top $659M 5 Sep, now $291M (56% off, 0.618 $252M, 0.66 $224M), RSI 4h 29. RH tiers all below 20d avg (BIG -37%, MID -56%, SMALL -39% off 30d highs); Solana tiers all above 20d avg, SMALL leading (+11% 14d vs BIG 0%). @insomniacxbt 2 Oct: on-chain sold first and will bottom first, AI value at $0.10.
- Raw (session scratchpad, temporary): macro/gt/*.json (RH, Solana, OKX 4h+1d), conf.py, confan.py, conf_rows.pkl, tiers.py, phase.py, sched.py, hold.py, order.py, sent2.py, xs_controlfull.json, pb/N_*.out.json, live_AI.json, live_PONS.json.

# v3, 3 Oct 2026 evening Bangkok: golden pocket per Ben, attention, size waves, phase framework
- [stated] Ben's golden pocket rule: zone 0.618 to 0.666 retracement of the run (0.5 = bounce zone, buy only in very bullish coins); wicks below are fine, invalid on a 1h close (or at least a 4h close) under 0.66; stop around a 0.75 pullback; target back to ATH. Possible add-ons: order block in the zone, RSI oversold (timeframe dependent), MA100/200 (1h/4h), buy pressure switch, holder count rising. Wants coin-level X sentiment history through every phase (PONS etc.), LunarCrush checked, size waves and parent coins rechecked on Solana and RH, and a phase framework he can apply to any new coin he sends.
- twitterapi.io key ran out of credits about 14:30 Bangkok 3 Oct 2026 ("Credits is not enough. Please recharge", HTTP 402 also from the droplet followwatch key, same key). Coin-level X history pull (28 cycle coins + WIF, PNUT, GOAT, MOODENG, ACT, FARTCOIN, CHILLGUY, PENGU, TRUMP, POPCAT; script macro/cxpull.py, 9,262 calls) is waiting on a recharge. Ben told by message.
- buyscore.py (buy-framework, md5 000e29249d5dbe2b9766af63cf930b85) uses an absolute QUIET_RATE of 0.045 contract posts/hour; twitterapi search density about doubled for every query from ~26 Sep 2026 (control query), so its "quiet" test reads too loud since then. Needs a control-normalised rate.
- Golden pocket test (4h candles, 49 coins RH/Solana GeckoTerminal + OKX 1h listed memes; run 3x+ inside 30 days, limit fill at 0.618, invalid on 4h close under 0.666, target ATH, 30 days, 2% costs): GP n=78: 19% reached ATH, 73% invalidated, avg +13% (survivor tailwind), median -13%, coin-balanced +5%. Same mechanics at 0.5 zone avg +11% (27% ATH), 0.38 zone +7% (34% ATH), 0.786 zone +27% (13% ATH). Hard stop at 0.75 instead: GP avg +14%, 22% ATH, 59% stopped. 1h-close invalidation: only 9 events (OKX listed coins rarely 3x in 30 days), all invalidated. Verdict: the pocket is an OK zone with a fat-tail payoff, not special vs other depths; deeper pays more per win.
- Add-ons at the pocket (n=78) made it WORSE: arriving with RSI 4h<35 (n=18) or RSI 1h<30 (n=14): 0% reached ATH, 83 to 93% invalidated; OB in zone (n=30) 7% ATH; MA100/200 in zone (n=12) 8%; 2+ add-ons (n=31) 3% ATH, avg -9%; NO add-ons (n=29) 28% ATH, avg +23%. Only a volume spike at the touch (2.5x, n=20) held level (+14 to +26%). Reading: a pocket reached slowly (not oversold) holds; a pocket reached in a crash keeps falling. Same pattern at 0.5 and 0.786 zones. TESTING, small n.
- LunarCrush coin topics (limited mode = last 30 days, enough for the RH cycle): $pons posts 8,347/day peak 6 Sep (day after the 5 Sep price top), engagement 10.5M -> 1.7 to 2.1k posts and 0.6 to 1.4M engagement 27 Sep to 3 Oct (posts -75%, engagement -90%); $moo posts peak 696 on 4 Sep, low 131 on 29 Sep (-81%) two days before the 1 Oct price low, then 350 on the 2 Oct bounce. LunarCrush sentiment % stays 72 to 97 and does not separate phases. "$ai" maps to a different coin (unusable for AI on RH).
- Size waves recheck: zigzag turn order of BIG vs MID vs SMALL tiers is not consistent on Solana (Dec 2025 to Oct 2026, few swings, window-edge noise); the RH Aug to Oct cycle fits Ben's order (pure memes and small stock memes topped 1 to 5 Sep, MOO/ASTEROID 10 Sep, MEME 13 Sep, AI 17 Sep, INU/SCHIFFY 19 to 20 Sep, then chain-wide decline; AI agent coins ORBIO/HARMONIC topped later, 1 Oct). Keep the state rule (big rolled over while small still rising = sell), drop any fixed lag.
- Phase framework for a new coin (v1, to apply on request): MARKUP (rising, within 30% of ATH, attention rising) -> TOP (near ATH, attention at its 30d peak, chain big coins rolling) = sell into strength -> FIRST DROP (20 to 50% off, attention still high) -> ANGER (50 to 80% off, attention 40%+ of its peak, angry words) = wait -> DEPRESSION (70%+ off, ATH 3+ days old, attention down 70%+ from peak, contract posts about 1/day or less) = buy zone (buy-framework STRONG BUY) -> RECOVERY (new 10-day high off the base, attention turning up). Pocket entries only on slow pullbacks; skip pockets hit in a crash (RSI 4h<35).
- Live read 3 Oct evening: PONS 56% off ATH, LunarCrush attention -75% posts / -90% engagement = late ANGER into DEPRESSION, price not yet 70% off. AI 66% off at the 0.66 pocket, arrived with RSI 4h 34 (crash-style arrival = weaker per the add-on test). MOO 84% off, attention low hit 29 Sep, bounced from 1 Oct = DEPRESSION to RECOVERY.
- Raw (session scratchpad, temporary): macro/gp.py, gp_rows.pkl, tiers2.py, h1/*.json (OKX 1h, 15 memes), cycle_coins.json, cxpull.py.
## v4 emotion framework (3 Oct 2026, cheat sheet stages labelled by LLM from X posts, tone not keywords)
- [stated] Ben asked: label X posts per coin into Wall Street cheat sheet stages, add talk volume as strength, test "all coins follow same human pattern", build counter-trade score, test on new coins, then use for baskets/chains; no alerts yet, build system first.
- Data: 50 coins (34 build, 16 holdout never used for tuning), ~4,900 coin-days, 4,003 labelled days + 2,114 market days (MKT_RH, MKT_MEMES, MKT_SOL). Files scratchpad macro/emo (ds.py, a1-a6.py, out_*.json).
- Same pattern? Half true. On average stages sit where the cheat sheet says: THRILL/EUPHORIA at the top (median 0-5% off top, loudest), ANXIETY/DENIAL 9-13 days after top (-60%), ANGER 50 days after (-73%), DEPRESSION 75 days after (-89%). But only 22% of coins pass through the stages in clean order; long-lived coins repeat mini cycles; HOPE shows both early and as bounce hope.
- RED (sell/scale out) = EUPHORIA, COMPLACENCY, THRILL, ANXIETY, loud DENIAL, or 3+ hot days (THRILL/EUPHORIA) in last 7, or cooling after heat. Frozen on build. Holdout: RED median 14d -19% vs 0% all days; 40% of RED days fell 30%+ in 14d vs 15% base; GREEN beat RED in 12 of 14 coins; RED lit within 5 days before to 3 after the real top on 88% of holdout coins. Still 10% of RED days went +100% in 14d, so scale out, not all out. Price-matched (same drawdown and trend) it still holds.
- GREEN (OPTIMISM, loud HOPE, quiet silence) is only a mild plus (+4% median 14d, 30% vs 25% chance of +50%). Not a strong buy signal.
- Bottom side: ANGER = still falling (build win 9%, 14d -10%). DEPRESSION slightly better than ANGER but too few cases to call. Sentiment does not time bottoms well.
- Macro: market-wide X mood (broad samples) is weak and not stable across halves. RH breadth of coin signals (share RED/HOT high = -24% next 14d; share dead/silent high = +26%) looked strong but only 46 days, one cycle, survivor coins. Solana breadth showed nothing. Not proven.
- Data trap (verified, see Pro Trading [[buy-framework]] v6): the 8x jump for coin queries from 26 Sep = X search showing the last ~7 days in full while older posts come back thinned (spam share 74% recent vs 8% older; real people only 1.5x; LunarCrush shows no jump). Not real crowd. Stage labels unaffected.
- [stated] Ben: rename RED/GREEN to SELL SOME / HOLD / NO SIGNAL; the cycle is fractal by chart timeframe (1m crowds cycle in minutes, daily crowds in weeks).

# v5 narrative visibility study, 3 Oct 2026 night Bangkok
- [stated] Ben: narratives (e.g. LONG eco "PvE, special liquidity") hit peak X visibility at the top; visibility rises exponentially (nobody, a few, more, everywhere), the top comes slightly after, then no new buyers. Same for chains (Robinhood chain hype; Solana fans saying bridge to Robinhood = near the RH top; "Robinhood killed Solana" talk). Believes Solana is now in a confident thrill phase and overextended; wants this linked to extension over MAs, VWAP, Supertrend.
- Method: 10 topics, 2 one-hour X windows per day, 1 Jun to 2 Oct 2026, counted on COMPARABLE posts (real people, 3+ likes, xsafe 1.7). Baskets: LONG (AI, BONER, MOO, SCHIFFY, ASTEROID, STRATEGY, SPACEHOOD, MEME), all 95 RH coins (median daily return), 15 Solana memes, SOL. Files scratchpad macro/vis (topics.json, topicvis.pkl, baskets.pkl, vis2 to vis5.py, ext.py). Cost about 1.3M X credits.
- RH wave 1: "robinhood chain" talk x96 in a week (1 to 7 Jul), peak 11 Jul (109 liked posts/h); "bridge to robinhood" peak 10 Jul; "Robinhood killed Solana" peak 13 Jul; RH coins kept rising to 21 Jul (10 days after the talk peak), then -32% by 2 Aug.
- LONG wave: LONG eco talk peak 3 Sep, PvE talk 2 to 3 Sep, PONS talk 1 to 5 Sep; LONG basket top 1 Sep (2 days before the talk peak), then -79% to 3 Oct. RH wave 2: "robinhood chain" talk peak 6 Sep, RH coins top 3 to 5 Sep, -23% since. General "solana" talk also peaked 7 Sep (crypto-wide attention peak).
- Reading: in all 4 waves the price top fell between 2 days BEFORE and 10 days AFTER the talk peak; talk growth went from x5 to x100 per week to under x1.5 at the top. Rule (TESTING, 4 events only): talk at a 60-day high after an explosive week, then weekly growth stalls under x1.5 = top zone, take profit; talk down 25% from its peak = the top is in.
- Coin level (50 coins, contract + ticker): price peak within -3 to +7 days of the talk peak on 43% of coins (median +2 days); many coins peak in talk at launch. Talk blow-off alone: 14d median -10 to -13% vs -4%; near the top with loud talk was NOT bearish (momentum). Emotion labels (SELL SOME) beat raw talk at coin level.
- Solana now (2 Oct): SOL near its 90d high (+46% since June) but general Solana talk is down to 34% of its 7 Sep peak, "solana is back" talk near zero, Solana memes talk at its low, LLM mood mostly hope/optimism (only 2 thrill days late Sep). Extension: SOL 15% over 50d avg, 39% over 200d. Not a crowd top by this data; pushback on "everyone agrees".
- Extension test (19 Solana/major coins, OKX daily, years): only EXTREME stretch helps: >100% over 50d avg -> next 30d median -22% (down 20% in 14d 22% vs 13%); >200% over 200d -> -18%; 20 to 60% stretch = no signal; 30d VWAP stretch same; Supertrend (10,3) flip down = no edge (-2% vs -2.5% base). Most stretched now: RAY (56% over 50d, 142% over 200d).

# v6 coin-level talk test + spam filter hygiene, 3 Oct 2026 ~22:30 Bangkok
- [stated] Ben asked: prove the talk-top rules on single Robinhood coins first (talk 2-month high + explosive week then growth under 1.5x; talk -25% from peak; new arrivals going exponential then stalling; big accounts converging), before any basket use. No narrative rules live yet. Keep the spam filter learning daily but never remove real posts.
- Test: 37 RH top coins, daily sample launch-1d to ATH+22d, comparable posts (real, 3+ likes, xsafe 1.9.1), daily volume estimated from page time span (95% of days had more than one page, so counts are estimates). Files scratchpad macro/vis/cdaily.json, cstudy.py, cstudy.pkl, cs_*.pkl.
- Result: talk 7d peak lands 0 to 7 days AFTER the price top on most coins (median +1 day; within 7 days on 25 of 37; before the top only 3 of those 25). New-arrival peak median 0 days, big-account peak +1 day. Coincident to lagging, not leading.
- Rule A (stall under 1.5x at a talk high) fired on only 7 of 37 coins; when it fired, 71% at or before the top, median 24% under ATH, but median 32% more upside after. Rule B (talk -25% from peak) fired 27 of 37, median 2 days after the top, median 62% under ATH: confirms late. Arrivals stall (N) fires too early in young coins (median 65% under ATH, +124% more upside).
- Talk divergence at new price highs (price new high, talk 25% under its 21d peak) = MORE upside (median +829% vs +281%), not a top. Bearish divergence fails on young coins.
- Verdict: on single RH coins this X data does not beat price for timing tops; basket/narrative level (4 waves) still unproven beyond 4 events. LLM emotion labels (v4 SELL SOME) remain the better coin-level tool.
- Spam filter: xsafe 1.9.1 live 3 Oct 22:05 Bangkok (md5 bce680b4694a088a6c60b3786bda4969, public copy same). Gold set 1,283 LLM-verified real posts: 1.8 removed 14 (about 1 to 2% of real posts: zero-width chars in emoji, coin lists, copytrading, leaderboard, long guides), 1.9.1 removes 0. Query-side exclusion terms A/B tested on 4 queries (654 real posts): 0 real posts lost.
- Hygiene process live on droplet: hook writes a daily kept/removed sample (/opt/common/xsafe/sample/), canary every 5 min (any gold post flagged = roll back to lastgood + one Telegram line), daily mine 08:13 Bangkok (copy-paste templates auto, regex ideas to review), drift vs frozen reference (target 99.5%+ real kept), scheduled task "X spam filter daily review" 08:41 Bangkok (needs "Require this computer" on to reach the droplet). The 14:56 UTC 3 Oct canary alert wrongly said "rolled back" (there was no safe copy yet); fixed.

# v7 100-coin talk test, real money, "it's dead", 3 Oct 2026 ~23:30 Bangkok
- [stated] Ben: believes talk extension over a moving average (like price over its MA) and new arrivals going exponential line up with price tops, and that social can be more precise than MAs/VWAP/Supertrend. Asked: 100 coins by market cap, three metrics (talk peak vs price peak; talk extension over MAs at hourly/4h/daily, find best MA; new people talking), plus the real money chain study and the "it's dead" study.
- Data: 100 coins pulled (51 RH, 24 Solana from GeckoTerminal peak mcap with a confirmed top 30%+ down; 24 OKX listed memes/alts with 2024 to 2026 tops), daily 1 page of comparable posts (real, 3+ likes, xsafe 1.9.1) ATH-30d to ATH+25d, post-top to 2 Oct every 2nd day, 4h windows ATH-7d..+3d for 15 big coins. 22 coins dropped for too little X talk (ANTFUN, BIRB, Ban, DONT, AOL, LetsBONK, TRUMP Solana, GOOSE, Hosico, astro ...), 78 used. Cost about 1.36M X credits (balance ~1.2M after). Files scratchpad macro/vis: universe2.pkl, cdaily3/4.json, c4h.json, cs2build.py, cs2.py, dead.py, h4.py, panel2.pkl; money/: m1.py, stable/tvl/dex/fees json, rh_blockscout.json.
- Data fix: gt price files for AI 0x2E8c31162b855A2ffa90F6F8634643Ad6F111e18 and PONS 0x39dbed3a2bd333467115de45665cc57f813c4571 hold each other's prices (file "AI" = PONS pool 0x10cc6bd38112cac182db90b6a71d8bb5939526ba). rh50.json ATH dates for AI/PONS are swapped (AI top 17 Sep, PONS 5 Sep). v6 coin numbers used the right PONS series.
- Talk peak (3d/7d avg) vs price top: median 0 days, within 7 days on 63%, before the top only 38%. Coincident, not leading.
- Talk extension (3d avg over 7/14/20/30d MA): peak lands within 3 days of the top on only 19 to 28% of coins (price extension 33 to 49%). Sell rule "extension went above 1.5 to 3x, now faded 30%": fires near the top (within 5 days) 19 to 30% of the time vs 15% for a random day; same rule on PRICE extension 50 to 67%. Higher talk extension raises 14d 30% drop odds only mildly (22% to 33 to 39%); price extension 9% to 46 to 52%. No MA length works for talk; 20d is least bad.
- New arrivals (first-time posters): peak a median 4 days BEFORE the price top (63% before; halves -4d and -2d), but a live rule (arrivals 25% under their peak while price within 15% of its high) hits the top no better than "price near its high" alone (24% vs 19% near top). Talk + price combo worse than price alone (42% vs 60%).
- 4h (12 coins, small n): talk peak median 24h before the 4h price top (54% fall 0 to 48h before vs ~20% random); price 5-day extension peak 69%. Price still as good or better.
- Verdict: on single coins X talk is coincident and less precise than price extension. Ben's "social beats technicals" thesis not supported at coin level. KILLED as a coin sell signal; 4h talk peak TESTING.
- Real money, Solana (2024 to Sep 2026, DefiLlama): big stablecoin inflow (30d > +8%) -> pump.fun basket next 21d -17% avg, 33% chance of another -30% vs 16% base (both halves); 7d inflow spike (z>1) negative both halves on memes and pump. Stablecoin flows show no pattern AT turns; DEX volume vs 60d avg about 1.25x at tops, 0.77 to 0.82x at bottoms. Low DEX activity is not a buy (memes -11% next 21d). No free Solana new-wallet source (bridges paid, no Dune key).
- Real money, Robinhood (one cycle): stablecoin inflow peaked the top week (+$195M week to 6 Sep), dropped to ~0 after 13 Sep; new wallets (173k/day) and active wallets peaked the week AFTER the top (13 Sep); DEX volume peaked the top week. Robinhood Blockscout charts via stats-service/api/v1/lines (Chrome only, Cloudflare blocks container and droplet).
- "It's dead" talk: posts with 3+ likes rarely say it (90% of 50%-down coin-days have zero); high share (3%+) -> 30d +33 to +90% and +50% odds 50 to 60%, but only 25 days on 6 to 8 coins. Quiet (talk 10% of its peak) -> another -30% only 3% vs 29% when still loud (6 coins). Not proven. "Dead" posters get few likes and X drops low-like posts after ~7 days, so history cannot be rebuilt; must be recorded live. Basket version not started (coin proof failed).

# v8 "it's everywhere" exposure study, 4 Oct 2026 ~00:40 Bangkok
- [stated] Ben: wants to measure peak exposure (the coin is unavoidable on crypto X, everyone agrees you must own it, the last buyer sees it), top in or shortly after; asked for ~50 approaches, influencer cohorts that newly pile in together, the best ones implemented; anger/scam/sold sentiment for bottoms; real money as a public doc for the risk on/off dashboard; whether new wallets add value or a paid source is worth it; 4h expansion (1h vs 4h). Zero-like posts not needed.
- Talk peak timing (77 coins) is two-humped: 27 to 30% of coins peak 8+ days BEFORE the top (launch hype, price climbs later on less talk), the rest cluster around the top (3-day talk peak within 2 days of the top 34%).
- Big-post pull: posts with 50+ likes (min_faves:50), 2-day windows, ATH-20d..+8d, 75 coins, with viewCount. Files scratchpad macro/vis cbig.json, expo.py, expo2.py, expo_panel.pkl, expo_R2.pkl.
- 16 measures x 2 live rules scored the same way (fire within 5 days of the top; fire 0 to 7 days BEFORE the top; price at fire vs ATH). Random day in the same window: 41% near, 27% before.
- WINNER: views of 50+ like posts, stretch = 3d avg / prior 14d avg (skip 3 days); rule "was 2x+ in the last 5 days, now 30% under that high" -> near top 56% (halves 55/59%), before or at top 48% (both halves 48%), price at fire median 26% under ATH, median 29% more upside after. New 100k+ follower accounts (KOL cohort) stretch: 64% near (61/68%), 49% before, but fires lower (41% under ATH). Big-post count and big-post likes: 53% near, 47% before. Influencer authors (20k+) fade: 55% near.
- Price stretch rule in the same window: 60% near but only 9% before the top, fires 41% under ATH (it confirms after). Raw talk (3+ likes) stretch 51% near / 34% before; reach, likes 49 to 53% near but only ~21% before; hype words, consensus (bull minus bear words), share of voice among RH coins, likes per post: no better than random.
- KOL cohort (new 100k+ accounts posting, 7d count) peaks in the 15 days before the top on 57% of coins, same day 7%, after 36%; adding it to the views rule did not improve it.
- Verdict: Ben's "exposure" thesis holds when exposure = VIEWS of big posts (not post counts): same hit rate as price stretch but fires before or at the top half the time and about 15 points higher in price. TESTING (75 coins, one 2026 cycle plus 2024 listed memes).
- Implemented: /opt/research/exposure/exposure.py on the droplet (public copy http://159.223.88.168/hx-c9ois/research/exposure.py, md5 a5d8ab690dbb342f6fbc7b753a0605f3): python3 exposure.py --ca <CA> --ticker <T> [--chain robinhood|solana] [--generic]; 30 days of 50+ like posts through xsafe, about 3k credits per coin; returns EXPOSURE PEAK PASSED = SELL SOME / EXPOSURE STRETCHED = top zone / NORMAL plus new 100k accounts in 7 days. 4 Oct 00:05 Bangkok: PONS views peaked 6 Sep (15.7M a day, top 5 Sep), both PONS and AI NORMAL now.
- 4h test, 19 coins (pull stopped by the burn brake, another session's xcollect3 was burning ~47k credits a minute): views per 4h too lumpy (one viral post dominates), signals near random; talk peak within 24h of the 4h top 58% (median 4h before) but live rules weak (best talk 168h fade 41% near vs 22% random). Answer: keep exposure at daily / 2-day resolution; 1h would be noisier and 4x cost.
- Anger / scam / sold / dead words (real 3+ like posts, coins 50%+ under top, 30 coins): pooled, high negative mood (6%+ of posts) -> another -30% only 5 to 8% vs 23 to 39% (both coin halves), but within the same coin it beat calm days on only 3 of 8 coins: a coin-type effect (older Solana memes chop, RH coins crash), not a timing signal. "Sold / took the loss" talk was worse (34% another -30%). Not proven.
- Real money public doc: http://159.223.88.168/hx-c9ois/research/real-money.html (md5 268a677e2b0674e28391a0b94d3e5780). Key: Solana stablecoins (7d avg) +8% in 14 days -> crash odds about 3x for 3 weeks (34 to 44% vs 9 to 15%) in 93% of 45 variants; pump.fun basket negative after inflows in every year; older memes positive in 2024 (bull fuel), negative 2025. RH: inflow peaked top week, wallets peaked a week after (lagging). Recommendation: no paid wallet source now; DefiLlama free per-app active users (pump.fun, PumpSwap, Raydium, fomo Wallet) only from 23 Aug 2026, record and test later. Readings 2 Oct: Solana +2.8% in 14 days (no flag), RH flat.
- X credits: Ben topped up 3 Oct night (~10.7M). Research spend this round about 0.7M.