How markets move, and what our AI actually looks at

Short version: there is no pattern that works for sure. Markets have small, unreliable tendencies. Traders who last find a slight edge, test it, and control risk so one wrong call cannot wipe them out.

Why there is no sure pattern

Edges get traded away. If a pattern reliably made money, large funds with more data and faster computers would trade it until it stopped working.
The crowd learns. What happened after a panic in one year may not repeat the next, because people remember and act differently.
Luck looks like skill. Test enough patterns and some will look great by pure chance. Real edges must survive on data they were not found on.

Crowd psychology: why prices overshoot

Herding. People buy because others are buying. Rising prices pull in more buyers (fear of missing out), so moves run further than the news justifies.
Panic and capitulation. Sharp drops trigger forced and fearful selling. Prices can fall below fair value, and the bounce afterwards is often strong but not guaranteed.
Overreaction and underreaction. Big surprises are often overdone for a few days. Slow-building news is often underpriced, which is one reason momentum (recent winners keep winning for months) has been one of the most persistent tendencies in research.
Anchoring. Traders fixate on round numbers and recent highs. Prices near a 52-week or 60-day high attract both sellers taking profit and buyers chasing breakouts.
Attention. Spikes in searches, news and social posts mark moments when the crowd is most emotional, at both tops and bottoms.

These are tendencies, right maybe 52 to 60 percent of the time, not rules. They work only across many trades with small position sizes.

What actually helps traders

Test first. Backtest an idea on real prices, then run it with practice money, then with small real money.
Control risk. Use stop losses and small positions. Most traders who fail lose on risk, not on picking.
Trade many times, not big once. A slight edge only shows up across many trades.
Distrust anything that sounds certain, including our AI.

Classic quant strategies, tested on 10 years of real daily prices

Return vs buy and hold, worst drop in brackets, 0.1% cost per trade. S&P 500 and Nasdaq futures, gold, Bitcoin, Ethereum.

Trend + momentum (above the 200-day average and MACD above zero; sell below the 200-day): S&P +135% vs +264% (−16%), Nasdaq +340% vs +544% (−21%), Gold +137% vs +230% (−31%), Bitcoin +8,245% vs +13,931% (−71%), Ethereum +2,114% vs +747% (−70%).
100-day EMA trend: S&P +77%, Nasdaq +185%, Gold +117%, Bitcoin +14,596% vs +13,931%, Ethereum +1,508% vs +747%.
200-day trend: S&P +96% (−20%), Nasdaq +252% (−22%), Gold +134% (−32%), Bitcoin +953% (−64%), Ethereum +2,005% (−71%).
Turtle breakout (buy a 20-day high, sell a 10-day low): S&P +81% (−17%), Nasdaq +146% (−21%), Gold +21% (−30%), Bitcoin +7,330% (−53%), Ethereum +744% (−51%).
RSI 2 dip in an uptrend (Connors): small gains with the smallest drops, S&P +23% (−13%), Nasdaq +42% (−11%).
Golden cross: S&P +96% (−34%), Nasdaq +291% (−28%), only 3 to 8 trades in 10 years.

Over a 10-year bull market, holding usually made more on stocks and gold. Trend rules mainly cut crashes, and beat holding on Ethereum and (100-day EMA) Bitcoin. Popular short-term ideas did worse: crossovers on 5-minute charts lost money after costs in public tests, and levels like the 200 EMA reacted no more often than a random line.

Verified on QuantConnect

We re-ran 5 trend strategies and buy and hold on QuantConnect, an independent professional backtesting platform. Daily data, $10,000, crypto from January 2018 and US funds from January 2010, to October 2026.

Ethereum (hold +127%, worst drop −93%): 200-day trend +1,509% (−72%), trend + momentum +1,035%, 100-day EMA +964%, golden cross +610%.
Bitcoin (hold +360%, worst drop −80%): 100-day EMA +1,407% (−56%), 200-day trend +743%, golden cross +613%, trend + momentum +510%.
S&P 500 (SPY), Nasdaq (QQQ), gold (GLD): buy and hold won every time (SPY +758%, QQQ +1,623%, GLD +235%). Trend rules there cut the worst drop roughly in half but made less.

Lesson: simple trend rules have paid off in crypto, where crashes are huge, and not in steadily rising stock indexes.

Options, volatility and global links

VIX seasonality (CBOE data, 1990 to today): fear is highest in October (average VIX 21.6) and March (20.5), lowest in July (17.7) and June (18.3). Markets tend to be calmest in early summer and most nervous in autumn.
VIX term structure: normally the 3-month VIX sits above the 1-month VIX. When the 1-month jumps above it, the crowd is panicking now. Our AI tests this ratio on every market.
Implied correlation (CBOE COR1M): how much options traders expect S&P stocks to move together. High means fear and herding, low means calm, stock-picking markets. It earned a vote for both Bitcoin and the S&P 500 in our tests.
Dark pools (SqueezeMetrics DIX and GEX): DIX estimates how much off-exchange trading was buying. GEX estimates options dealers' gamma: high gamma tends to calm markets, negative gamma tends to amplify moves. GEX earned a vote for the S&P 500.
Max pain: the option strike where option buyers would lose the most at expiry. Prices sometimes drift toward it into expiry, but it is a weak tendency, not a magnet. Briefs show it from Deribit for Bitcoin, Ethereum, Solana, XRP, Avalanche, Tron and Hyperliquid (Deribit has no stock options).
European vs American options: European options (SPX, most index options, Deribit crypto) can only be exercised at expiry and settle in cash. American options (SPY, single stocks) can be exercised any day, which matters around dividends and makes early assignment possible when you sell them.
Korea and the US: over the last 10 years, KOSPI and S&P 500 daily returns had a correlation of 0.39, and weekly returns 0.59. They often move together, so KOSPI is one of the cross-market signals the AI tests.

Microstructure and quant toolkit

Order flow imbalance: taker buys minus taker sells, plus resting bids minus asks near the price. Short-term pressure, not a long-term signal.
Micro-price: the mid price weighted by the size waiting at the best bid and ask. It leans toward where the next tick is more likely to go.
Kyle's lambda: how much price moves per dollar of net buying. Higher lambda means a thinner, more easily pushed market.
VPIN: flow toxicity over equal-volume buckets. Spikes often come before volatile moves.
Hawkes process: trades excite more trades. We estimate how much of the activity is self-triggered from how bursty trade arrivals are.
Meta-orders and optimal execution: big players split one large order into many small ones to limit impact. Using lambda, briefs estimate the cost of a $10k, $100k and $1M order and how many slices it would need.
Queue position: on an exchange your limit order waits behind earlier orders at the same price. This needs full order-by-order data, which is not free, so we do not model it.
Cointegration and stat arb: some pairs (Solana/Ethereum, EUR/GBP) drift apart and snap back. The AI tab tests 8 classic pairs and flags when the spread is 2 standard deviations wide.

How our AI works

For each market, the AI measures how well every data source below came before that market's moves over the next 10 days (we tested 5, 10 and 20 days on real data; 10 worked best across markets), across all the history we have. A signal only gets a vote if its record is statistically meaningful (t-statistic of at least 2) and held up in both the older and the newer half of the history. If nothing passes, the AI makes no call. Every call is logged and graded 10 days later, and the AI's own strategy is backtested using only data that was available at the time.

The crowd psychology section counts what each market did after panic days, euphoria days, 5-day streaks, capitulation (20 percent below its recent high) and new highs, compared with an ordinary day.

Every data source we use (all free and public)

Prices: Coinbase (crypto), Yahoo Finance (futures, forex, indexes), Alpaca (US stocks).
Crowd mood: Crypto Fear and Greed Index (alternative.me), StockTwits bullish share, ApeWisdom Reddit mentions, Hacker News stories (Algolia).
Attention and news: Wikipedia daily page views, GDELT global news tone.
Market fear and macro: VIX and other series from the St. Louis Fed (FRED).
Economic calendar: Forex Factory weekly calendar of high and medium impact releases.
Big players: SEC Form 4 insider filings, CFTC Commitments of Traders, FINRA short volume.
Washington (free Quiver-style data): government contracts from USAspending.gov and lobbying filings from the Senate LDA (lda.gov), for US stocks.
Derivatives and on-chain: OKX funding rates and long/short ratios, open interest, Deribit implied volatility, Coin Metrics on-chain data, blockchain hashrate, DeFi TVL, stablecoin supply.
Prediction markets: Polymarket and Kalshi.
Cross-market moves: S&P 500 futures, US dollar, 10-year Treasury, gold, oil and Bitcoin, 5-day changes.
Fed data (FRED): yield curve, Fed funds rate, dollar index, financial stress, financial conditions, Fed balance sheet, inflation expectations, oil, jobless claims, reverse repo, Treasury cash.
More Fed data: oil and gold volatility indexes (OVX, GVZ), Nasdaq volatility (VXN), high-yield and Baa credit spreads, US crude inventories, gasoline price, 10-year and 3-month yields, Michigan consumer sentiment, euro-dollar rate, plus the copper-to-gold ratio.
Options and dark pools: SqueezeMetrics DIX and GEX, CBOE COR1M implied correlation, CBOE VIX and VIX3M term structure, Deribit options open interest for max pain, and Korea KOSPI.
Levels and order flow: prior day, week, 20-day and 52-week highs and lows, moving averages and round numbers; Coinbase trade tape for buyer vs seller aggression; Deribit options gamma per strike for call walls, put walls and the gamma flip.
Calendar: day-of-week and month effects, shown only when statistically significant.
Optional, with a key: Grok summaries of posts on X.

Sources change and can go offline. When one is missing, the AI simply works with fewer signals.

Educational content, not investment advice. Past tendencies do not guarantee future results. See our Risk Disclosure.