Falak (@falaktb)

£25,000 quant setup feat 196GB VRAM High frequency tick data means processing thousands of price updates per second across multiple instruments. With 196GB of VRAM, the entire system stays in memory. Full transformer weights plus live context windows per symbol, so inference runs continuously without waiting on data transfer. 64 CPU cores handle parallel preprocessing. Tick data requires normalisation, feature extraction, and sequencing before it reaches the model. When hundreds of symbols stream at the same time, single threaded pipelines fall behind instantly. 256GB of ECC RAM because a single bit flip corrupting position sizing or risk logic is never worth the savings. DDR5 bandwidth keeps the GPUs fed and latency stable when loading batches at scale. Quick walkthrough of the components, then a live demo running the model on raw tick data. #quant #machinelearning #nvidia #blackwell #workstation

Gaussian distributions are symmetrical. Options markets aren't. By taking the second derivative of call prices against the strike (the Breeden Litzenberger extraction), we map the exact risk neutral probability density. The resulting 3D cone shows you the exact probability mass of future price action. Normal markets don't skew. That distortion is the literal footprint of heavy institutional hedging. Stop guessing what the smart money is doing. #derivatives #quant #londonstrategicedge #options #apple

Fine tuning a 20 billion parameter model on 75,000 domain specific examples. The base model starts generic. Every financial question returns surface level retail content. The goal is to shift billions of weights until every response reflects institutional grade quantitative analysis. We load the full 20B parameters across dual GPUs. LoRA adapters sit inside the attention blocks, updating a controlled fraction of the weight matrix. 190GB of VRAM holds batch sizes of 15, which most setups cannot sustain. 75,000 instruction response pairs. One full epoch. Three and a half hours of gradient descent. Before training: vague, generalised output. After training: statistical reasoning, drawdown mechanics, volatility regime context. No more retail noise. #quantfinance #machinelearning #finetuning #llm #deeplearning

Today, we’re announcing the LSE Terminal. A new foundation for quantitative trading.

A random forest is a collection of decision trees that vote on an outcome. Each tree trains on a slightly different sample of the data, reducing the chance that any one tree dominates the result. #randomforest

Principal component analysis reduces a set of correlated assets down to the few underlying factors that explain most of their movement. It exposes how concentrated a portfolio really is, since positions that look separate often trace back to the same factor. It is the first thing to run when you want to see what is truly moving a portfolio. #PrincipalComponentAnalysis #quant

#quant #londonstrategicedge

#opensource

(Construction going on outside)* #h100 #tcn #alternativedata #machinelearning #quantfinance

A Kalman filter keeps an estimate of a hidden value together with the variance of that estimate, which is the minimum variance solution under Gaussian process and measurement noise. Each step updates the estimate with a model, raises the variance by the process noise, then revises the estimate by the Kalman gain multiplied by the difference between the measurement and the prediction. #quantfinance #kalmanfilter #timeseries #statisticalmodeling #machinelearning

Processing market data in hardware rather than software is the single variable that gives FPGA firms a structural edge at the matching engine. #fpga #hft #quant #optiver

Part 2 will cover how we minimise latency through colocation and FPGAs. #predictionmarkets #polymarket #hft #quant #arbitrage

Level 3 order book data captures every order hitting the matching engine and every update that follows, while Level 1 only shows what actually trades. That gap is exactly why quant desks pay for direct exchange feeds. Once you can see where liquidity is actually resting, the next price move becomes a function of the book state rather than a guess. Order book imbalance and queue position both live in Level 3 and simply don’t exist in Level 1. #hft #algotrading #orderflow #microstructure #wallstreet

Pearson correlation measures linear dependence between two return streams as a single number between minus one and plus one. Compute it across thousands of instruments and you get a correlation matrix, the structural map of how the market moves together. Pairs that share a factor (same currency leg, same sector, same risk regime) cluster near plus or minus one by construction. The residual between them, fit by OLS, is the variance the shared factor cannot explain. If that residual is stationary, it mean reverts. That is the trade. When the spread deviates beyond two sigma, you sell the side that moved and buy the side that didn't, then exit at the mean. This is the math behind statistical arbitrage, the strategy that has generated billions in P&L for hedge funds since the 1980s. #quantfinance #algotrading #pearsoncorrelation #hedgefund #correlation

Most of the work in financial machine learning is getting the data into a usable state. If the timestamps are off or the series are inconsistent the research stops being reliable and the model output stops meaning much. #WalkForwardValidation #CleanData #MachineLearning #Quant #QuantFinance

HFT economics are driven by scale. If the edge on each trade is small but the number of opportunities is large, lowering latency becomes worth an enormous amount of money. #HFT #QuantFinance #AlgorithmicTrading #MarketMicrostructure #Trading

Every quant interview gets to conditional probability at some point. In financial markets, probabilities are not fixed. New information arrives and the conditional distribution changes. A large part of quantitative finance is estimating those shifts correctly. That is also why Hidden Markov Models matter. The regime is not directly observed. You estimate the probability of being in each hidden state from observed data, then update that estimate as new observations arrive. #quant #bayestheorem #cardcounting #statistics #quantfinance

Bid and ask are only the first line of the order book. Level 3 data records every order added, cancelled and modified in the book. From that you can reconstruct queue position, liquidity imbalance and how depth shifts across price levels through time. That is the layer most quantitative execution and market microstructure models operate on. #quant #orderbook #level3data #marketmicrostructure #orderflow

The prop firm model depends on traders not understanding the maths. #quant #quantfinance #riskmanagement #statistics #propfirms

Self hosted ML infrastructure: clean datasets, built in features, hyperparameter tuning, 30 model templates. Open access for users to train and tune#machinelearning #quant #nvidia #selfhosted

Standard Monte Carlo averages risk across history. It treats a calm month and a stressed month as interchangeable samples. They are not. Regime conditioned paths reprice tail risk based on what state you are in now and how long that state typically lasts. #QuantFinance #RiskManagement #MonteCarlo #MarkovChains #SystematicTrading

How quants manage risk. Position sizing is the variable most traders ignore. On institutional desks it is the foundation. Exposure scales with edge. When signal quality is weak, size stays small. When the edge proves itself through time, size increases mechanically. Not conviction. Expected value and drawdown control. #quantfinance #riskmanagement #machinelearning #dataanalytics

Inside a neural network. Not the theory. The actual model. 8 inputs go in. Each layer compresses the information. Learns relationships no human would code manually. By the output, you get a probability. Not a guess. A confidence score. #quant #machinelearning #neuralnetwork #deeplearning