现在跟2000年的不同之处是花尔街预测水平高了几个数量级

[noktard]

如果2000年大家能预测到今天msft能涨到500，03年就不会跌到20以下了。

Wall Street uses AI in two main ways: prediction and risk management. Despite the hype, the real edge often comes from the second.

1. Prediction (with AI/ML models)

Firms try to forecast returns, volatility, and risks — but markets are noisy, so predictions aren’t perfect.

Price/return forecasting

Deep learning (LSTMs, transformers) used to detect patterns in time series data.

Example: predicting short-term price moves using order book data.

Limitation: models can “overfit” — markets evolve faster than models adapt.

Earnings & fundamentals

NLP models analyze earnings calls, filings, news sentiment.

Transformers (like BERT or GPT variants) score management tone and forward-looking guidance.

Helps anticipate surprises before consensus changes.

2.Alternative data
Risk Management (where AI really shines)[/b]

This is often more valuable than prediction.

Volatility forecasting

ML models detect stress build-up (e.g., unusual derivatives activity).

Helps firms hedge before volatility spikes.

Liquidity monitoring

AI watches market depth in real time — signals if liquidity is drying up (potential flash crash).

Fraud detection & compliance

Models catch unusual trading behavior, spoofing, or insider-like activity.

Portfolio optimization

Reinforcement learning is used to rebalance portfolios dynamically under risk constraints.

Satellite images (retail parking lots, oil storage tanks).

Credit card transaction data.

Web scraping consumer reviews.
AI models integrate this with financial data to predict revenue or demand.

What AI Has Changed

Markets are faster to react — fewer delayed crashes like 2008, more quick corrections (COVID 2020).

Prediction is still limited: even the best AI can’t reliably forecast “black swan” events (pandemics, wars, political shocks).

Risk control is much better: firms can cut exposure quickly, preventing collapses from spreading as widely.

In short: AI hasn’t made Wall Street able to “see the future,” but it has made it much better at reacting to stress before it snowballs into a systemic crash. That’s why crashes look sharper but shorter today.

[noktard]

1. Two Sigma Investments

How they use AI:

Heavy users of machine learning and deep learning for market prediction.

They ingest massive alternative data sets — satellite imagery, shipping data, weather patterns, web activity.

NLP models (including transformers) analyze corporate filings, earnings calls, and global news.

Goal: Find tiny predictive signals (“alpha”) that humans would never notice, then scale them across thousands of securities.

2. Renaissance Technologies

Renaissance (Jim Simons’ fund, famous for Medallion Fund) is very secretive.

Publicly known: they rely on statistical and ML models trained on decades of historical data.

While they were early pioneers in non-linear models, they are rumored to have experimented with neural networks for decades.

They don’t rely much on “macro predictions” — instead they exploit short-term inefficiencies.

3. Citadel

How they use AI:

AI models to manage market-making and liquidity provision.

Real-time order book prediction with reinforcement learning.

Deep learning models to predict short-term volatility and price moves in equities and options.

Impact: Citadel Securities is one of the largest market makers in the world, and their AI-driven systems let them provide liquidity (and profit from spreads) while hedging risk efficiently.

4. Point72 (Steve Cohen’s fund)

They built an in-house unit called Point72 Cubist, focusing on ML-driven strategies.

Uses transformer-based NLP for sentiment analysis on earnings calls, news, and even CEO interviews.

AI also helps in portfolio construction — dynamically rebalancing based on risk scenarios.

5. Smaller, newer players (like Numerai)

Numerai runs a crowdsourced hedge fund where data scientists worldwide build ML models (often deep learning).

Uses meta-models to combine predictions from thousands of participants.

Heavy use of transformers for time-series and text.

[noktard]

1. Probabilistic models are much better now

2000s: Quant models were mostly statistical regressions and some early neural nets, limited by computing power and data availability.

Today: Hedge funds use deep learning, transformers, reinforcement learning, and massive alternative datasets (satellite images, credit card data, NLP on earnings calls).

This means they can spot dislocations faster and arbitrage them away before prices get wildly misaligned.

2. Policy responses are faster

In 2000–2008, central banks often moved slowly (Fed took years to cut rates in the dot-com bust, months to stabilize banks in 2008).

In 2020 (COVID crash), the Fed pumped trillions in within weeks. That’s why the S&P dropped 34% but recovered within 5 months, instead of dragging on for years.

Faster intervention prevents drawn-out 80% collapses in blue chips.

3. More passive + algorithmic investing

Index funds and ETFs create a constant bid for large-cap stocks.

High-frequency traders and market makers provide liquidity that dampens extreme crashes.

Together, this means stocks don’t get “abandoned” the way Microsoft was left for dead in 2003.

4. But efficiency ≠ no crashes

You still see sharp drawdowns:

Meta lost ₇₅% from 2021 to 2022.

Many SPACs and unprofitable tech stocks fell 80–90%.

The difference: mega-cap leaders (MSFT, Apple, Amazon, Nvidia) don’t fall that far anymore, because too many sophisticated models and institutions are watching them.

Result: You rarely get the “Microsoft at 80% off” opportunity in quality large caps.

But inefficiencies still exist — just in riskier corners (small caps, emerging markets, or during sector panics).