Лабораторный журнал: 100+ гипотез, большинство уже мертво.

After volume-confirmed dump events, mid-frequency reflexive bounce trades on 30m/1h/2h/4h timeframes carry sufficient edge for systematic deployment with tight SL/TP.

Deep multi-timeframe path-dependent backtest on 90 days of 5min klines across 78 symbols (T1+T2+T3). Realistic filters: BTC regime gate, time-of-day, 60min cooldown. Per-tier breakdown.

Initial test on n=100 random sample claimed +16.8%/mo. Bootstrap CI on that sample was ±0.70%, mean +0.41% — statistically indistinguishable from zero. Multiplying noise by 30× trade frequency compounded into a fake-positive projection.

Mean falls inside zero band. No statistically significant edge.

Borrowing from Perelman's reduced-length functional from Ricci flow theory, "smooth" price approaches to a signal trigger (low geometric path energy) should correlate with higher follow-through than chaotic approaches.

Discretized for log-price path on 240×1m bars before signal trigger:

Where τ_k = k/N is normalized time position, σ²_k is local realized variance in 10-bar window centered on k, and r_k is the log-return at bar k.

The internal scoring function of a legacy strategy (V14 QUANT, stored in 14.6M signalsnapshots) may be overfit. Higher quality scores may not correlate positively with realized P&L.

Random sample 5000 snapshots. For each: fetch forward 1m klines, simulate entry at the snapshot's own entryTop/Bot zone, walk forward bar-by-bar checking SL/TP/timeout. Compute realized pnlPct per trade. Then rank-correlate features vs outcome.

Following Sornette's Log-Periodic Power Law model, financial bubbles exhibit faster-than-exponential growth with log-periodic oscillations preceding a critical time t_c. We test whether crypto's accelerated bubble cycles (days–weeks rather than months) yield tractable t_c predictions on BTC, ETH, and top alts.

Parameter ranges from Sornette's published constraints:

• m ∈ [0.1, 0.9] — critical exponent (typically 0.33)
• ω ∈ [6, 13] — log-periodic frequency (typically 6.36)
• tc — critical time (singularity), constrained to next 1–14 days
• A, B, C, φ — fit parameters via Levenberg-Marquardt

1. Rolling LPPL fit on 1h log-prices, window 30–90 days, refit every hour
2. Validity gates: 0.1<m<0.9, 6<ω<13, R²>0.9, t_c within 1–14 days forward
3. Parameter stability check across multiple window shifts
4. Multi-asset confluence (3+ synchronous bubble signals = strong)
5. Entry: SHORT setup 2–5 days before t_c
6. Exit: at t_c or on LPPL fit breakdown

• Sornette (1996–present), Financial Crisis Observatory, ETH Zurich
• 27+ peer-reviewed papers documenting fits on major historical bubbles (1929, 1987, 2000, 2007)
• Mixed crypto results: BTC 2017 and 2021 peaks fit well ex-post; out-of-sample mixed

Binance MONITORING-tag announcements (volatility-warning labels) are preceded by detectable volume + range anomalies in the 6h window before the public announcement. If true, a pre-event SHORT detector could capture the post-announcement drop before the crowd reads the news.

Phase 0: scrape Binance CMS catalog (catalogId=49). 20 pages × 50 articles. For each: identify coin, check if it has Binance USDT-M futures listing, fetch 1m klines spanning [T-6h, T+1h] around announcement. Compute pre-event drift, volume ratio, range expansion.

100% downward drift in pre-event window is suggestive but n=3 is noise. Bootstrap 95% CI on n=3 spans entire ±50% range.

• Most MONITORING tags target spot-only altcoins, not futures-tradable
• CMS pagination rate-limited beyond ~16 articles
• Older announcement pages removed from CMS archive

Top-PnL Hyperliquid traders publish every position on-chain (HL's order book is fully transparent). Mirroring their entries to Binance USDT-M futures should capture a fraction of their edge — particularly on majors (BTC, ETH, SOL) where slippage is minimal.

Identify top-20 wallets by 30-day PnL on HL. WebSocket listen to their position changes. When a tracked wallet opens a position above $100K notional, fire a mirroring order on Binance same coin, same side.

1. Latency mismatch: HL whale fills are typically maker-limit at deep liquidity. Any follower with public-API latency enters 0.3-1.5% worse than the originator on average.
2. Position duration mismatch: Whales hold hours-to-days with average DCA-in over multiple fills. Our single-shot mirror catches only the entry tip, exits poorly.
3. Wallet attribution: Same person operates multiple wallets. Composite NET exposure ≠ individual wallet signal.
4. Selection bias: "Top 30-day PnL" includes survivorship + recency bias. Wallets fall off the leaderboard the moment they have a drawdown.
5. Reverse-MEV risk: When public HL whale wallets are watched, they sometimes intentionally fake entries to trap copy-traders.

Binance DELISTING announcements (catalogId=48) trigger immediate panic dumps. Auto-shorting the announced coin within seconds of announcement should capture −15 to −40% over 24-72h.

• Massive gap-down on first fill: by the time futures opens enough liquidity to short, price has already dumped 20-40%. Average entry slippage 8-15% from announcement price.
• Spot dump > futures dump: spot market panics first; futures follows but at different magnitude due to existing long liquidations forcing a bounce.
• Squeeze risk: after initial dump, surviving longs squeeze shorts on thin liquidity. 30-40% counter-bounces common within 24h.
• Liquidity collapse: trading volume on delisted coins evaporates within 12h, making exit difficult without significant slippage.

MONITORING tags (catalogId=49, BURST-class strategies) are different — they happen pre-announcement with intact liquidity. DELISTING auto-short specifically refers to acting AFTER public announcement, which has structural reasons to fail.

Aggregating signal feeds from multiple "smart money" sources (HL alerts, on-chain whale flows, exchange leaderboards) into a unified scoring system should produce above-baseline performance through diversification of signal sources.

• Signal aggregation amplifies false signals: when 3 independent feeds say BUY simultaneously, often it's because they're all reading the same already-developing pump that's about to reverse
• No regime filter: equally weighted signals in trending vs choppy regimes — chop regime saw 90% of losses
• Implicit correlation between "smart money" sources: HL whales, Twitter influencers, and on-chain trackers all watch each other. Independent signals are not actually independent.

Sub-second mean reversion on tick data is profitable for sufficiently fast operators.

• Colocation in exchange datacenter required (~$5–10K/mo per venue)
• FPGA / kernel-bypass networking for sub-100μs latency
• FIX or proprietary binary protocols, not REST
• Capital floor for meaningful PnL at 0.5–2 bps/trade is enterprise-tier, not retail
• Compliance overhead: KYC, AML, market-making agreements

Pivoting to mid-frequency (1–5min hold) as accessible alternative for the broader retail/prop universe.

BTC and ETH log-prices are cointegrated; the stationary spread mean-reverts, so trading the z-score of the residual yields market-neutral edge independent of direction.

Step 1 — estimate the hedge ratio $\beta$ by OLS on log-prices:

Step 2 — test the residual for stationarity (ADF). Trade the z-score:

Rolling 90-day window, refit $\beta$ daily, ADF gate at $p<0.05$. Path-dependent replay on 1h bars, funding + 6 bps round-trip fee applied to both legs. 2024–2026 out-of-sample.

• The hedge ratio $\beta$ is non-stationary — ETH/BTC structurally re-rates across regimes, so yesterday’s spread is not today’s spread.
• Cointegration that holds in calm regimes breaks exactly when it matters (trend onset), producing the worst losses when $|z|$ is largest.
• Two-leg funding + fees eat the thin residual; the edge is real gross, negative net.

A market-neutral basket of correlated alts has a residual that follows an Ornstein–Uhlenbeck process; trading deviations beyond a calibrated band captures reversion with positive expectancy.

Model the de-meaned basket spread $X_t$ as:

Mean-reversion speed $\theta$ gives the half-life, which sets holding time:

Optimal entry band scales with the stationary standard deviation $\sigma/\sqrt{2\theta}$.

Estimate $\theta,\mu,\sigma$ by MLE on a rolling window. Enter at $\pm1.5$ stationary-sigma, exit at $\mu$, hard time-stop at $3\times t_{1/2}$. 30 liquid USDT-M alts, path-dependent.

The cointegration failure of static pairs (see N-018) is fixable: let the hedge ratio $\beta_t$ evolve as a hidden state estimated online with a Kalman filter, so the spread stays stationary as the relationship drifts.

State (hidden hedge ratio) and observation (one leg priced off the other):

Kalman update for the posterior estimate:

Process/observation noise $Q,R$ tuned by likelihood. Trade the standardized innovation $y_t-\hat\beta_{t-1}x_t$. Prototype only — not yet replayed on full out-of-sample with realistic costs.

Decompose the alt return covariance into principal components; the top eigenportfolios are systematic risk, the residual (idiosyncratic) returns mean-revert. Trade the residual, hedged against the top factors.

Eigendecompose the correlation matrix $C = V\Lambda V^\top$, keep top-$k$ factors $F$, regress each asset on them:

Trade the cumulative idiosyncratic residual as an OU process (cf. Avellaneda & Lee 2010):

Top-5 eigenportfolios as factors on 40 alts, daily refit, dollar-neutral residual book, 1h replay with full two-sided fees + funding + borrow.

The Hurst exponent $H$ distinguishes trending ($H>0.5$) from mean-reverting ($H<0.5$) regimes. Gating an existing momentum book to fire only when $H>0.55$ should cut whipsaw losses.

Estimate $H$ from the scaling of the rescaled range $R/S$ with window length $n$:

Fit $\log(R/S)$ vs $\log n$; the slope is $H$. $H=0.5$ is a random walk.

Compute $H$ on rolling 256-bar log-returns (DFA cross-check). Use as a binary gate on a baseline EMA-cross momentum book across 50 symbols.

Position size inversely proportional to forecast volatility (vol-targeting) improves risk-adjusted return versus fixed size, because realized vol is forecastable even when returns are not.

Conditional variance recursion:

Scale exposure to a target volatility $\sigma^\*$:

Fit GARCH(1,1) per symbol by QMLE, forecast 1-step $\hat\sigma_t$, scale a fixed momentum book to 60%/yr target vol. Compare to flat sizing.

The Cont–Kukanov order-flow imbalance at the best bid/ask linearly predicts the next mid-price move; a fast book-pressure signal captures sub-minute drift.

OFI aggregates signed size changes at the top of book over $[t-\Delta,t]$:

Predict the mid-move linearly:

Reconstruct L1 book from the public depth stream, regress next-$\delta$ mid-move on OFI for $\delta\in\{1,5,15\}$s. Net-of-fee replay assuming taker entry.

Easley–López-de-Prado VPIN (volume-synchronized probability of informed trading) spikes before volatility events; fading extreme VPIN, or stepping aside, avoids adverse selection.

Over $n$ equal-volume buckets, VPIN is the average absolute buy/sell imbalance:

Bulk-volume classification (BVC) on aggregate trades, 50-bucket VPIN. Tested as a volatility pre-warning and as a fade trigger. Blocked by classification noise on public tape.

Quoting symmetric bid/ask around a reservation price with inventory-aware skew, per Avellaneda–Stoikov, earns the spread on liquid alts while controlling inventory risk.

Inventory-adjusted reservation price ($q$ = inventory, $\gamma$ = risk aversion):

Optimal half-spread combines risk and book-arrival intensity $\kappa$:

Paper market-maker on 3 liquid alts, inventory cap, $\gamma$ and $\kappa$ calibrated from the arrival curve. Maker rebate modeled, latency penalty applied on requotes.

• We get filled preferentially on the wrong side — informed flow lifts our stale quote right before the move (adverse selection), exactly the cost AS assumes you can offset with requote speed.
• At public-API requote latency we cannot pull quotes fast enough; faster co-located MMs are always ahead in the queue.
• On trends the inventory term cannot keep up; the book runs against the held inventory.

Trade arrivals cluster (one trade raises the probability of the next), modelled by a self-exciting Hawkes process. The fitted branching ratio may flag the build-up of a liquidation cascade before price confirms.

Intensity rises with each past event and decays exponentially:

The branching ratio $n=\alpha/\beta$ measures endogeneity; $n\to1$ signals criticality.

Plan: fit $\mu,\alpha,\beta$ online by MLE on the aggregate-trade stream per symbol; monitor $n=\alpha/\beta$ as an early-warning of self-reflexive flow. Not yet built.

When the perpetual funding rate is persistently positive, a delta-neutral position (short perp / long spot) harvests funding with no price exposure — a pure carry trade.

Delta-neutral funding accrual over holding horizon (8h intervals):

Annualized carry from a flat funding rate $f$ paid 3×/day:

Rank symbols by trailing funding, hold delta-neutral while $f>$ threshold, unwind on flip. Real spot+perp fees, borrow, and spot-leg slippage applied.

The perpetual trades at a basis to spot that must converge via funding; entering when the basis is wide and holding to convergence is a low-risk trade.

Basis is the perp–spot gap, pinned by the funding mechanism:

Enter delta-neutral when $|b_t|$ exceeds a threshold, exit on convergence. Sub-second snapshots across 3 venues, full fees.

The 25-delta risk reversal (call IV minus put IV) encodes directional positioning; extreme skew is a contrarian signal for the underlying.

Skew as the IV gap between symmetric out-of-money strikes:

Plan: pull Deribit BTC/ETH option surfaces, build the 25Δ RR term structure, test extremes as contrarian spot signals. Blocked: liquid crypto options exist only for BTC/ETH, so the cross-section is two names wide.

The MVRV Z-score (deviation of market cap from realized cap) marks macro tops and bottoms; extreme readings time multi-month BTC cycle turns.

Standardize the gap between market value and realized value:

Historical MVRV-Z on BTC since 2015. Label $Z>7$ as distribution, $Z<0$ as accumulation. Evaluate forward 90/180-day returns from threshold crossings.

A 3-state Gaussian HMM on returns recovers latent bull / chop / bear regimes; conditioning exposure on the most-likely state improves return.

Emissions are state-conditional Gaussians; states evolve by a transition matrix $A$:

Fit by Baum–Welch (EM); decode the state path by Viterbi.

Fit 3-state HMM on rolling returns, go long only in the inferred high-mean state. Walk-forward, costs applied.

Price loops across three pairs (e.g. BTC→ETH→USDT→BTC) drift out of parity; executing the cycle when the product of rates exceeds 1 locks a riskless profit.

Arbitrage exists when the product of the three exchange rates beats fees:

where $\phi$ is the per-leg taker fee.

Stream all three books, compute $R$ continuously, simulate atomic 3-leg execution with real taker fees and realistic fill latency.

A discrete wavelet transform separates price into time–frequency bands; trading the denoised low-frequency trend while ignoring high-frequency noise improves momentum timing.

Decompose the log-price into approximation + detail coefficients across levels $j$:

Reconstruct using only low-frequency approximation bands as the "trend".

Daubechies-4 DWT, trade the slope of the reconstructed trend. Tested with both batch and strictly-causal (online) transforms.

EMD adaptively decomposes price into intrinsic mode functions (IMFs); the residual + low-order IMFs form a clean trend tradeable without lag.

Sifting extracts each IMF as the signal minus the mean of its upper/lower envelopes:

EMD sifting on rolling windows; trade trend = residual + lowest-frequency IMF. Causal vs batch comparison.

Markets carry a measurable dominant cycle (Ehlers); estimating its period and phase times entries at cycle troughs and exits at peaks.

Pick the period maximizing spectral power over a rolling window:

Rolling DFT / Hilbert dominant-cycle on 1h bars across 30 symbols, enter on inferred cycle phase, costs applied.

SSA decomposes a series into trend, oscillatory, and noise components via the eigentriples of its trajectory matrix; the reconstructed trend is a lower-lag filter than moving averages.

Embed the series into an $L\times K$ trajectory matrix and SVD it:

Group leading eigentriples as "trend", reconstruct by diagonal averaging.

Causal SSA on a sliding window; the open question is whether grouped reconstruction lags less than an EMA of equal smoothness, net of recompute cost.

A recurrent LSTM trained on multivariate features (OHLCV, OI, funding) predicts the next-bar return well enough to trade directionally.

Minimizing MSE on returns drives the output to the conditional mean:

2-layer LSTM, 40 features, walk-forward train/validate, predict next 1h return; trade sign of prediction with costs.

Self-attention captures long-range dependencies a recurrent net misses; a Transformer should beat the LSTM (N-038) on directional forecasting.

Attention re-weights value vectors by query–key similarity:

Encoder-only Transformer, same feature set and walk-forward protocol as N-038, tuned over depth/heads.

Gradient-boosted trees on engineered features (vol regime, funding, OI delta, time-of-day) can classify which signals of an existing book are higher-quality, used as a filter rather than a generator.

The model is an additive ensemble fit greedily to the gradient of the loss:

Label baseline signals win/loss, train XGBoost with monotonic constraints, evaluate as a probability gate on out-of-sample signals. Strict temporal split, no leakage.

A PPO agent learns to slice a parent order to minimize implementation shortfall versus TWAP/VWAP, adapting to live book conditions.

PPO maximizes a clipped policy ratio against the advantage estimate:

Train in a simulated LOB with a market-impact model; reward = negative implementation shortfall. Evaluated in-sim only.

Evolving populations of indicator-threshold rules with a genetic algorithm discovers profitable combinations no human would hand-craft.

Fitness-proportional selection over generations:

GA over indicator/threshold/logic genes, fitness = in-sample Sharpe, 200 generations, then hold-out test.

BOCPD (Adams–MacKay) flags structural breaks in the return-generating process in near-real-time; cutting risk at a detected changepoint avoids regime-transition losses.

Recursively update the posterior over the current run length $r_t$:

BOCPD with a Gaussian observation model and hazard $1/\lambda$ on volatility-standardized returns; de-risk when changepoint probability spikes.

A GP with a tuned kernel gives calibrated predictive means and variances for returns, letting us trade only high-confidence (low-variance) forecasts.

Posterior mean and variance at a test point:

GP with RBF + periodic kernels on feature windows; trade only when predictive variance is low. Costs applied.

Large net inflows of a coin to exchange wallets precede selling; net outflows precede accumulation. Trading the netflow z-score anticipates pressure.

Standardize daily net exchange flow:

On-chain exchange-flow series for majors, z-score extremes as directional bias, forward 1–3 day returns. Daily granularity only.

When SOPR crosses 1 (aggregate coins moving at a profit vs loss), it marks capitulation/euphoria inflection points tradeable on swing horizons.

Ratio of realized value to value at creation, over spent outputs:

BTC SOPR (and short-term-holder SOPR) vs forward swing returns around the 1.0 cross.

A low SSR (large stablecoin supply relative to BTC market cap) means high "dry powder" and is bullish; trading SSR extremes anticipates rallies.

SSR and its Bollinger-band oscillator vs forward 30-day BTC returns since 2019.

The Network-Value-to-Transactions ratio (market cap / on-chain transaction volume) is crypto’s P/E; high NVT = overvalued = short.

NVT and NVT-signal (smoothed) vs forward returns; controlled for the structural decline in on-chain volume as activity moved to L2s/exchanges.

The Puell Multiple (daily miner revenue vs its yearly average) marks cyclical extremes via miner capitulation/euphoria.

BTC Puell vs forward 90/180-day returns across cycles; halving-aware.

Rising search interest in a coin precedes retail inflows and price; trading the search-volume z-score front-runs the crowd.

Weekly/daily Google Trends SVI for top coins vs forward returns; lead-lag cross-correlation.

Aggregate sentiment polarity from crypto X posts (transformer classifier) predicts short-horizon returns.

FinBERT-style polarity on filtered cashtag streams, hourly aggregate vs next-hour return. Bot/spam filtering attempted.

The composite Fear & Greed index at extremes ("Extreme Fear" / "Extreme Greed") is a contrarian signal for multi-day reversals.

Daily FNG since 2018; forward 7/14-day BTC returns conditioned on tail buckets.

The Higuchi/box-counting fractal dimension $D$ of a price window distinguishes smooth trends ($D\to1$) from rough chop ($D\to2$); regime-gating on $D$ improves a trend book.

Higuchi FD on rolling 128-bar windows as a trend/chop gate, 30 symbols, costs applied.

A positive largest Lyapunov exponent $\lambda_1$ means sensitive dependence (chaos, low short-term predictability); $\lambda_1\to0$ windows are more forecastable.

Average exponential divergence of nearby trajectories in the embedded phase space:

Rosenstein algorithm on delay-embedded returns; gate forecasting models to low-$\lambda_1$ windows.

The Shannon entropy of the binned return distribution measures disorder; low-entropy windows (concentrated returns) carry more exploitable structure.

Entropy of binned rolling returns; correlate with subsequent strategy hit-rate and with volatility-of-volatility.

RQA metrics (determinism, laminarity) on the recurrence plot of a price window detect impending regime transitions before variance reacts.

Delay-embed returns, build the recurrence matrix, track determinism/laminarity as early-warning. Spot-checked, not fully replayed.

Genuine trade-size and volume first digits follow Benford’s law; venues or pairs with manufactured volume deviate, flagging wash-traded markets to avoid.

Chi-square / MAD of first-digit distribution of trade sizes per pair-venue vs Benford; cross-check against known-clean majors.

Most eigenvalues of an empirical correlation matrix are noise (within the Marchenko–Pastur band); cleaning them yields a more stable correlation estimate for risk and hedging.

Eigenvalues below $\lambda_+$ are statistically indistinguishable from noise, where:

Clip/shrink eigenvalues inside the MP band, rebuild $C$, use for minimum-variance hedge weights vs raw sample $C$.

Transfer entropy from coin A to coin B measures directed information flow; trading laggards when leaders move captures a predictable spillover.

Estimate pairwise TE across 40 coins, trade detected laggards on leader moves, costs + latency applied.

Persistence diagrams of a sliding point-cloud of returns capture topological "holes" whose birth/death (persistence landscapes) precede crashes (Gidea–Katz).

Track features across filtration scale $\epsilon$; persistence is death minus birth:

$L^p$ norm of the persistence landscape as a crash early-warning.

Plan: sliding-window point clouds of (multi-asset) returns, Vietoris–Rips filtration, monitor landscape norm. Not yet built.

Price regime shifts are cusp catastrophes; fitting the cusp normal form to a control/state embedding predicts sudden jumps.

Cobb stochastic-cusp fit (control variables = vol, OI) to BTC; likelihood-ratio vs linear model.

BTC trades inversely to the US dollar index; using DXY moves as a directional overlay for BTC adds edge.

Rolling 30/90-day BTC–DXY correlation; trade BTC on DXY signal when correlation is strongly negative.

BTC behaves as digital gold; gold momentum leads BTC, exploitable as a cross-asset overlay.

Lead-lag cross-correlation BTC vs XAU since 2017.

BTC’s beta to Nasdaq (NDX) spikes in risk-off shocks; using NDX futures overnight as a risk gate avoids gap-down sessions.

Condition BTC exposure on overnight NDX-futures direction during US-correlated regimes; flat otherwise.

The ETH/BTC ratio trends in multi-week regimes; rotating between ETH and BTC by the ratio’s momentum beats holding either.

Dual-EMA crossover on ETH/BTC, rotate the held leg, weekly rebalance, costs applied; 2019–2026.

Certain weekdays (e.g. "Monday pump", "weekend drift") carry persistent directional bias exploitable systematically.

Per-weekday mean returns with block bootstrap CIs; rolling-window stability check across years.

Volatility has a robust intraday seasonality (Asian / EU / US session overlaps); sizing and stop distances should adapt to the hour.

Average absolute return by UTC hour (vol seasonality):

Estimate the intraday vol curve per symbol; scale position size and ATR-stops by the hour-of-day factor.

Full-moon / new-moon phases (and other calendar folklore) correlate with crypto reversals.

Returns bucketed by lunar phase, full-moon windows tested for reversal bias; false-discovery-rate controlled across many calendar hypotheses.

Price impact is linear in signed order flow with slope $\lambda$ (Kyle); estimating $\lambda$ per symbol lets us cap order size so impact stays below expected edge.

Mid-price moves linearly in net signed volume:

Cap participation so expected impact $<$ expected edge:

Estimate $\lambda$ by regressing mid-moves on signed volume per symbol/hour; size orders to keep modeled impact under a fraction of expected edge.

The microprice — mid adjusted by book imbalance — is a better short-horizon fair-value estimate than the mid, improving passive fill timing.

Weight the two sides by the opposite queue size:

Compute microprice from L1, test as next-tick mid predictor and as a passive-quote anchor.

Order flow is long-memory — trade signs are positively autocorrelated for hours (Lillo–Farmer) — so recent net flow predicts continued same-direction pressure.

The sign autocorrelation decays as a slow power law:

Estimate sign-ACF, trade continuation on persistent net flow, costs + latency applied.

Transaction prices bounce between bid and ask, creating predictable short-term reversion (Roll); fading the bounce harvests it.

Roll: observed price changes have negative first-order autocovariance from the bounce:

Fade last-trade-side at micro-horizon; full taker fees applied.

Hidden iceberg orders reveal themselves through repeated same-size refills at a level; detecting them locates strong support/resistance to lean on.

Reconstruct per-level refill patterns from L2 updates; tag persistent hidden liquidity. Limited by public depth-stream snapshotting.

Large fleeting orders that cancel before execution (spoofs) signal the spoofer’s true intent; fading the displayed pressure captures the real move.

Flag rapid place-then-cancel clusters; trade opposite the fake pressure.

Less-liquid alts (high Amihud illiquidity) carry a return premium for bearing liquidity risk; tilting toward them harvests it.

Rank alts by Amihud, long the illiquid tertile / short the liquid, monthly rebalance, costs (which are themselves higher for illiquid names).

The effective spread can be backed out from the serial covariance of price changes (Roll), giving a transaction-cost estimate without quote data.

Estimate Roll spread per pair, compare to observed quoted spread, use to flag costly markets for the backtester.

One venue leads price discovery (e.g. the highest-volume perp); its moves predict laggard venues by tens-to-hundreds of milliseconds.

Synchronized multi-venue tape, cross-correlation lag estimate, trade laggard on leader move.

AMM (DEX) prices lag CEX during fast moves; arbitraging the dislocation captures the convergence.

AMM marginal price from constant-product reserves:

Monitor top-liquidity pools vs CEX mid; simulate atomic arb net of gas, swap fee, and bridge/settlement risk.

Aggregate dealer gamma flips the market between mean-reverting (positive GEX) and trend-amplifying (negative GEX) regimes; GEX sign sets the playbook.

Reconstruct GEX from the Deribit options chain; condition a mean-reversion vs momentum book on GEX sign.

Implied variance exceeds realized variance on average (the VRP); systematically selling options/variance harvests the premium.

Delta-hedged short-vol / short-strangle on BTC-ETH, measure VRP capture vs tail losses.

Extreme put/call volume ratios are contrarian sentiment signals for the underlying.

BTC/ETH option PCR extremes vs forward returns.

The sign and level of (IV − RV) signals whether to be net long or short volatility/gamma across the book.

Track DVOL-style IV vs realized vol; use as a regime tilt for vol-sensitive sleeves (not as a standalone options trade).

Dated quarterly futures trade at a basis to spot that decays to zero at expiry; a calendar (long spot / short quarterly) harvests the roll-down.

Annualized basis from time-to-expiry $\tau$:

Long spot / short quarterly when annualized basis exceeds a threshold, hold toward expiry, roll. Real fees + collateral cost.

The same perp can have positive funding on one venue and negative on another; going long the negative-funding venue and short the positive-funding venue nets the spread, delta-neutral.

Monitor funding across venues, open offsetting legs when the spread exceeds cost, manage collateral on both sides.

Temporary stablecoin depegs (e.g. USDC to $0.97) revert to $1.00; buying the depeg captures the reversion.

Buy below a peg threshold, size by an estimated permanent-failure probability, exit at reconvergence. Studied on historical depeg events.

Sizing each trade by a fraction of the Kelly-optimal bet maximizes long-run log-growth without the ruinous variance of full Kelly.

Optimal fraction from edge $b$ and win prob $p$:

We deploy a de-rated fraction $\kappa f^\*$ with $\kappa\approx 0.25$–$0.5$.

Apply quarter-to-half Kelly using empirically estimated (not assumed) edge/odds per strategy; compare growth and drawdown vs fixed-fractional.

Allocating capital across uncorrelated strategy sleeves by equal risk contribution (rather than equal capital) improves the aggregate Sharpe.

Choose weights so each sleeve contributes equal risk:

Estimate the sleeve covariance (cleaned via N-058 RMT), solve for equal-risk-contribution weights, rebalance monthly, compare aggregate Sharpe vs equal-capital.

A basket of 3+ correlated alts has a stationary linear combination (Johansen cointegrating vector) more robust than a single pair (N-018).

Test the rank $r$ of the cointegration space via the trace statistic:

Estimate cointegrating vectors on rolling 90d baskets, trade the stationary combination z-score, full multi-leg costs.

A copula captures non-linear, tail-dependent co-movement that linear cointegration misses; trading conditional-probability extremes of the copula beats z-score pairs.

Fit Student-t / Clayton copulas to pair return ranks, trade mispricing index extremes, costs applied.

The classic Gatev distance method — pair assets by minimal normalized-price distance, trade 2-sigma divergences — works on crypto majors.

Form pairs on a 12-month formation window, trade 6-month, normalized prices, costs applied; replicated across 2019–2026.

Bullish/bearish divergence between price and the Relative Strength Index precedes reversals.

Algorithmic divergence detection (price extreme vs RSI extreme), forward-return test across 50 symbols, costs applied — removing the human eye that usually selects only the winners.

The Ichimoku Kinko Hyo system (cloud, conversion/base lines, lagging span) generates profitable trend signals.

Full Ichimoku ruleset (cloud breakouts + TK cross + Chikou confirm) backtested across 40 symbols and parameter variants.

Price moves in 5-3 Elliott wave patterns with Fibonacci-ratio retracements (0.382, 0.618) that can be counted and traded.

Attempted to encode objective wave-counting rules; tested Fibonacci retracement-level bounces as entries across symbols.

W.D. Gann’s geometric angles (1×1, 2×1) and "squaring price and time" identify support/resistance and turning points.

Tested Gann-angle support/resistance and time-squares with mechanized rules on multiple unit scalings.

A convolutional net trained on rendered candlestick chart images learns profitable visual patterns end-to-end.

Render OHLC windows to images, train a CNN classifier on forward-return labels, strict temporal split.

An autoencoder trained on "normal" market states flags anomalous regimes (high reconstruction error) to step aside before disorderly moves.

Train on calm-regime feature vectors, monitor reconstruction error live, de-risk on spikes.

Clustering feature vectors (vol, trend, volume, dispersion) into $k$ market states lets each strategy run only in its favourable cluster.

K-means / GMM on standardized regime features; map historical strategy performance per cluster.

Stacking many weak predictors under a meta-learner yields a strong combined return forecast.

Stack LSTM, XGBoost, logistic, GP base learners under a meta-regressor with nested cross-validation.

Near large monthly option expiries, spot gravitates toward the "max-pain" strike (where most options expire worthless), tradeable in the final hours.

Deribit BTC/ETH monthly expiries, measure spot drift toward max-pain in the final 24–48h vs control days.

Recurring inflows around month-end/start (DCA buyers, payday) create a turn-of-month bullish drift.

Turn-of-month window returns vs rest-of-month, block bootstrap, stability across years.

Returns or volatility carry long memory (fractional integration $d$); an ARFIMA model exploits it for forecasting.

Discretize returns into states (e.g. quintiles) and use the empirical transition matrix to forecast the next state’s expected return.

Estimate the transition matrix on rolling windows, trade the expected next-state return, costs applied.

Assets with more negative realized skew earn higher subsequent returns (compensation for crash risk); a skew-sorted tilt harvests it.

Rank alts by trailing realized skew, long most-negative-skew / short most-positive, monthly rebalance, costs applied.

The Heterogeneous Autoregressive model of realized volatility (daily/weekly/monthly RV components) forecasts volatility well enough to drive sizing and stop placement.

Fit HAR-RV on intraday realized variance per symbol, forecast next-day RV, drive vol-targeted sizing and ATR-stop scaling; benchmark vs GARCH (N-023).

An asset’s own past 1–3 month return predicts its next-month return (Moskowitz–Ooi–Pedersen); going long recent winners / short recent losers across crypto harvests trend.

Position = sign of the past-$k$ return, vol-scaled:

Lookbacks $k\in\{20,40,60\}$d, vol-targeted, weekly rebalance over 30+ liquid alts, costs and funding applied, 2019–2026 walk-forward.

Ranking alts by trailing return and going long the top decile / short the bottom decile captures a cross-sectional momentum premium, market-neutral.

Long the top-decile, short the bottom-decile by trailing return $r_{i,t-k:t}$:

Decile sort on 30d trailing return, dollar-neutral long/short, weekly rebalance, borrow + fees applied; tested against the well-known "momentum crash" episodes.

Institutions ("smart money") leave a footprint — order blocks, liquidity pools resting above/below swing highs and lows, and fair value gaps (FVG). Price returns to these zones to fill imbalance and grab liquidity; trading the zones front-runs institutional flow.

The only objectively-codable pieces. A 3-candle fair value gap (bullish):

A liquidity sweep / stop-run = wick beyond a prior swing extreme, close back inside:

Mechanized the falsifiable rules — liquidity sweep of a prior swing, then entry on return into the nearest FVG — and tested forward returns across 50 symbols with full costs, removing the human who normally draws the zones after the fact.

The original "smart money" theory: a Composite Operator accumulates in trading ranges, leaving a schematic (selling climax, spring, sign of strength, last point of support, phases A–E). Entering on the "spring" — a false breakdown of the range low that rapidly reclaims — front-runs the markup.

Spring = penetration of the range low reclaimed within $n$ bars on elevated volume:

Codified only the spring (the rest of the schematic is labeled in hindsight). Tested spring entries inside identified ranges, forward returns, costs applied.

Price touching the lower Bollinger Band (20, 2σ) is "oversold" and mean-reverts to the mid-band — the entry taught in virtually every beginner course.

Bands are a rolling z-score envelope:

The taught trade: long the touch, exit at $\mu_{20}$.

Long on close back above the band after a touch, exit at mid-band, −3% failsafe, 12h cap. 30 most-liquid USDT-M perps, 84 days of 5m bars, 0.20% round-trip friction. n = 11,938 trades.

• A 2σ touch on a trending asset is not "oversold" — it is the trend working. The band is a volatility envelope, not a value signal.
• The gross edge is a few basis points at best; 20 bps of friction per round trip turns a coin-flip into a wood-chipper.
• t = −19.9 across twelve thousand trades. This is not "needs better filters". This is dead at any confidence level you like.

When the fast EMA crosses above the slow EMA the trend has turned; ride it until the death cross. The most-viewed strategy format on trading YouTube.

EMA_t = \alpha C_t + (1-\alpha)\,EMA_{t-1}, \qquad \alpha=\tfrac{2}{n+1}

The cross is a lagging filter of a lagging filter — by construction it fires after the move.

Long on 9>21 cross, exit on reverse cross or 48h. Same universe / friction / split as N-109. n = 9,373.

• On 5m crypto the cross fires dozens of times per week per symbol; four of five entries are chop, and the one real trend does not pay for the other four after fees.
• The classic defence — "use it on higher timeframes" — reduces n so far that survivorship and window luck dominate any course-seller's screenshot.

RSI below 30 marks capitulation; buying the cross back above 30 captures the bounce. Sold as a "high win-rate" entry in countless courses.

RSI = 100 - \frac{100}{1+RS}, \qquad RS=\frac{\overline{gain}_{14}}{\overline{loss}_{14}}

Long when RSI crosses up through 30, exit at RSI 50 or 12h. Same universe / friction / split. n = 4,012.

• The 48% win rate looks respectable — and is exactly the trap. Winners are small (bounce to RSI 50), losers are the trades where "oversold" kept being right.
• Crypto downtrends routinely pin RSI below 30 for hours; the strategy systematically buys the middle of liquidation cascades.

A close above the recent swing high ("break of structure" in SMC vocabulary) signals continuation; buy the breakout. The trend-following half of every SMC course.

\text{BOS}: \quad C_t > \max(H_{t-48..t-2}) \times 1.002

Long the breakout close, SL −2%, TP +4%, 12h cap. Same universe / friction. n = 2,284, time-split 60/40.

• Crypto 4h-scale swing breaks are where breakout sellers live: the market-maker fades the obvious level, and the retail stop cluster above the high is the exit liquidity.
• Negative in both time halves — not a regime artifact.

When price wicks below a swing low ("sweeps liquidity") and closes back above, smart money has filled longs on retail stops; buying the reclaim rides the reversal. The single most-marketed SMC setup.

\text{sweep}: \quad L_t < 0.997\\times\\min(L_{t-48..t-4})\\ \\wedge\\ C_t > \min(L_{t-48..t-4})

Long the reclaim close, SL under the sweep wick, TP at 2R, 8h cap. Same universe / friction / split. n = 1,496.

• The mechanically-defined sweep-and-reclaim is indistinguishable from a downtrend pausing. One wick below a low carries no information about who traded it.
• It tested worse than random entries with the same R:R — the setup actively selects weak tape.
• The narrative ("smart money grabbed liquidity") is unfalsifiable dressing on a losing pattern. Nobody selling the course shows you n=1,496.

A three-candle gap (candle 1 high below candle 3 low) is an "imbalance" that price returns to fill; buying inside the gap rides the continuation.

\text{bullish FVG}: \quad L_{t} > H_{t-2}\times 1.006, \qquad \text{entry at } \tfrac{L_t + H_{t-2}}{2}

Limit at gap midpoint on retrace, SL under gap base, TP 2R. Same universe / friction / split. n = 1,151.

• There is a faint gross tendency for gaps to act as support — and it is almost exactly the size of retail friction.
• t = 1.1 on eleven hundred trades means the observed +0.05% is indistinguishable from luck.

After an impulse leg, the 0.618–0.65 retracement zone is a high-probability continuation entry — the "golden pocket" of fib-based courses.

\text{zone} = [\,H - 0.65\,(H-L),\; H - 0.618\,(H-L)\,], \qquad \text{impulse } \tfrac{H-L}{L} \ge 6\%

Long the first touch of the zone (setup void on new high or 0.786 break), SL under 0.786, TP at the prior high. Same universe / friction / split. n = 462.

• Strip the fib mysticism and this is buying a pullback inside a fresh impulse trend — a documented momentum effect. The 0.618 level itself is likely incidental; the conditioning on impulse is the substance.
• It beats the random-in-uptrend control by ~1pp/trade, so the timing adds real value beyond trend context.
• But the test half kept only a third of the train edge — consistent with regime sensitivity, thin n, or slow decay.

The last opposing candle before an impulsive move ("order block") marks institutional inventory; price returning to that zone finds support. The most defensible claim in the SMC canon.

\text{impulse}: \tfrac{C_t - C_{t-3}}{C_{t-3}} \ge 2.5\%, \qquad \text{OB} = \text{last red candle in } [t-10, t-3]

Entry on first return into $[L_{OB}, H_{OB}]$; SL under the block; TP at 2R.

Same universe / friction / split as the whole series. n = 604. Benchmarked against BOTH controls (random 2:1: −0.17%; random-in-uptrend: −0.18%).

• Formalized, an order block is a pullback-to-origin entry inside a fresh impulse — again the momentum-pullback effect, with the zone anchored at the move's origin where real prior transactions cluster.
• It beats both random controls by a full percentage point per trade with t = 5. This is the only course-sold setup in our audit with a statistically solid, both-halves-positive result.
• No claim about "institutions" is needed or supported — the mechanism is consistent with plain anchoring + momentum.

Meta-entry. Any long entry rule with SL −1.5% / TP +3% evaluated in our window should be benchmarked against firing the same trade at RANDOM times: if a "strategy" cannot beat random, its content is zero.

Expected net of a random 2:1 long with win probability $p$:

Random timing in our window realizes $p=0.38$ → $\mathbb{E}[r]=-0.17\%$: the friction floor.

5,517 pseudo-random long entries (fixed grid) + 1,713 random entries conditioned on an active uptrend, identical exits, identical friction, same 30-symbol universe and 84 days as the whole N-109…N-116 series.

• Six of eight setups sold as "top strategies" in $500–$2,000 courses are indistinguishable from or worse than random trading after realistic fees.
• The two survivors (order block, golden pocket) are re-brandings of the oldest documented effect in the book — momentum pullback — and their marketing never mentions the mechanical conditions under which they test positive.
• Win-rate screenshots, discretionary "confirmation", and unfalsifiable smart-money narratives are how a −0.2%/trade pattern is packaged as a livelihood.

The MACD histogram turning positive signals momentum ignition; long the cross, exit the reverse cross. Arguably the most-screenshotted indicator in existence.

MACD_t = EMA_{12}(C) - EMA_{26}(C), \qquad hist_t = MACD_t - EMA_9(MACD)

Long on histogram cross-up, exit on cross-down or 48h. 30 most-liquid USDT-M perps, 84d of 5m bars, 0.20% round-trip friction, time-split. n = 13,430.

• MACD is a smoothed difference of smoothed averages — twice-lagged price. On 5m crypto it fires constantly and catches every chop.
• Negative in both time halves (−0.15/−0.19). Thirteen thousand trades leave no room for "you used it wrong".

Tenkan crossing above Kijun with price above Kijun is the canonical Ichimoku long trigger — sold as a complete "system" in weekend courses.

Tenkan = \tfrac{\max H_9 + \min L_9}{2}, \qquad Kijun = \tfrac{\max H_{26} + \min L_{26}}{2}

Long TK cross-up with price above Kijun, exit reverse cross or 48h. Same harness as the whole series. n = 6,785.

• Donchian-midline crosses are just slower price lags with Japanese names. Same disease as MACD: late entry, chop churn, fees.
• The mystique (cloud, five lines, 1930s rice-trader lore) sells courses; the math is two moving midpoints.

A candle with a long lower wick (≥2× body) printed at a local low shows "rejection" — buyers stepped in; long the close. Price-action course staple #1.

wick_{low} \ge 2\times|C-O| \;\wedge\; wick_{low} \ge 0.6\,(H-L) \;\wedge\; L_t \le \min L_{48}\times1.002

Long the pin close, SL under the wick, TP 2R, 8h cap. Same harness. n = 5,866.

• On a 5m chart a "rejection wick" at the low is usually a liquidation ripple inside a continuing move, not accumulation.
• PF 0.56 across six thousand trades, negative in both halves. Among the worst risk-adjusted results in this entire log.

A green candle whose body engulfs the prior red body at a local low signals reversal — long the close. Price-action staple #2.

C_{t-1}O_t \;\wedge\; O_t \le C_{t-1} \;\wedge\; C_t \ge O_{t-1}

Long the engulfing close near a 48-bar low, SL under both lows, TP 2R. Same harness. n = 11,303.

• Eleven thousand trades, t = −16.6. The pattern fires everywhere in crypto chop and pays the spread every time.

"Institutions execute at VWAP", so price pulling back to the daily VWAP from above finds algorithmic support — long the touch. A day-trading course favorite.

VWAP_t = \frac{\sum_{session} C_i V_i}{\sum_{session} V_i}

Price above VWAP for 1h, touch and close back above → long, SL −1%, TP +2%. Daily anchor (00:00 UTC). Same harness. n = 2,592.

• Execution desks track VWAP to minimize slippage — they do not defend it. There is no bid wall living at the average.
• Negative in both halves; equal to the random-entry floor. The touch carries zero information.

A false break below a mature trading range that instantly reclaims ("spring") marks the end of accumulation — the classic Wyckoff long trigger, heavily marketed in 2024-26 course culture.

\text{range}_{96}: \tfrac{H_r-L_r}{L_r}\le 5\% ; \quad \text{spring}: L_t \in (0.99\,L_r,\,L_r), \; C_t > L_r

Long the reclaim close, SL under the spring wick, TP 2R, 12h cap. Same harness. n = 2,058.

• Mechanically, a spring is indistinguishable from a range starting to fail — and on crypto perps ranges resolve down as often as up.
• Worse than the random floor. Same verdict as its SMC twin, the liquidity sweep (N-113): the reclaim carries no directional information.

"Volume confirms the move": a strong green candle on ≥4× average volume signals institutional participation — long the continuation.

V_t \ge 4\,\overline{V}_{72} \;\wedge\; \tfrac{C_t-O_t}{O_t}\ge 0.4\%

Long the spike close, SL −1.5%, TP +3%, 8h cap. Same harness. n = 1,988.

• A 4× volume candle on 5m crypto is usually the climax, not the beginning — you are buying the moment the aggressor finishes.
• Matches the random floor exactly: the volume filter adds zero information to the breakout it "confirms".

A tight consolidation ("base") followed by an impulsive rally marks a demand zone; the first return to the base finds unfilled buy orders. The supply-and-demand course canon.

\text{base}: \tfrac{H_b-L_b}{L_b}\le 1\% \text{ over } \ge 30\text{min}; \quad \text{impulse}: +3\% \text{ in } 30\text{min}; \quad \text{entry at first return to } H_b

Long the first retest of the base top, SL under the base, TP 2R. Same harness. n = 59 (the setup is genuinely rare when defined strictly).

• Strictly defined, this is once again a pullback to the origin of a fresh impulse — the same effect that keeps Order Block (N-116) and Golden Pocket (N-115) alive. Three brandings, one phenomenon.
• n = 59 is thin. The direction and both-halves consistency match its siblings, which is why we grade it partial rather than noise.