Guides

Backtest a value-at-risk model: count breaches, run the Kupiec frequency test and the Christoffersen independence test, and read clustered breaches.

8-minute read

Build an efficient frontier: estimate expected returns and the covariance matrix, solve minimum-variance portfolios, and handle estimation error.

8-minute read

Calibrate probability forecasts: score with Brier and log loss, read a reliability diagram, apply Platt or isotonic recalibration, confirm the fix.

8-minute read

Choose a risk-adjusted return metric by matching what it penalizes to the risk you care about: Sharpe, Sortino, Calmar, Omega, information ratio.

8-minute read

Choose between Sharpe and Sortino by matching the metric to your returns: Sharpe for symmetric returns, Sortino when downside is the concern.

7-minute read

Estimate drawdown recovery time from a strategy's return moments: run a Cornish-Fisher Monte Carlo, read the recovery percentiles, plan for the tail.

8-minute read

Read the options Greeks in plain terms: delta, gamma, theta, vega, and rho, what each measures, how they interact, and how to manage a position.

8-minute read

Size positions with the Kelly criterion when your edge is uncertain: estimate edge and variance, apply fractional Kelly, and stress-test the drawdown.

9-minute read

Avoid backtest overfitting with a practical workflow: limit trials, hold out data, measure the probability of overfitting, prefer robust rules.

8-minute read

Compute a deflated Sharpe ratio step by step: gather the inputs, count trials, adjust for skew and kurtosis, and read the probability the edge is real.

8-minute read

Generate synthetic market data for testing: choose a process that captures fat tails and volatility clustering, then use it without overfitting.

8-minute read

Run walk-forward validation: choose anchored or rolling windows, set fit and test lengths, re-optimize per window, and aggregate out-of-sample results.

8-minute read

Test a pair for cointegration with Engle-Granger: estimate the hedge ratio, run an ADF test on the residual, measure the half-life, validate it.

9-minute read

A step-by-step process for validating a trading strategy without fooling yourself: out-of-sample testing, trial accounting, deflated Sharpe, and a VaR backtest.

8-minute read

Backtest with realistic slippage: model the spread, temporary and permanent market impact, latency, and partial fills so paper returns survive live.

9-minute read

Choose a market data vendor by matching coverage, resolution, and history to your strategy, then comparing total cost of ownership across vendors.

9-minute read

Audit an LLM research prompt for look-ahead leakage: scan for outcome-revealing prices, directional words, and future facts the model could cheat with.

8-minute read

Build a RAG pipeline over SEC filings: ingest and chunk 10-Ks, embed and retrieve passages, ground answers with citations, and verify extracted numbers.

10-minute read

Build a regression suite for a finance prompt: collect representative and adversarial cases, define expected outputs, and run it on every change.

8-minute read

Choose between RAG and fine-tuning for filings: match the method to whether the task needs fresh knowledge or a fixed format, then weigh the cost.

9-minute read

Cut LLM token cost in a finance agent: right-size the model, cache the stable prefix, trim context, batch deferrable work, measure cost per decision.

9-minute read

Decide between batch and real-time LLM calls by deadline: route latency-tolerant jobs to batch for the discount, keep waiting-on work real-time.

7-minute read

Defend a finance LLM against prompt injection: treat retrieved content as untrusted, separate instructions from data, restrict tools, test attacks.

8-minute read

A practical guide to putting a large language model into a finance workflow safely: scoping the task, grounding answers, guarding inputs, and verifying every number.

9-minute read

Design a fallback chain for LLM providers: order models by quality and cost, set timeouts and retries, handle rate limits, and simulate failures first.

8-minute read

Detect hallucinations in finance LLM output: verify every number against the source, check citation faithfulness, and flag unsupported claims.

8-minute read

Estimate the cost of an AI research agent: count tokens per step, multiply by tool calls and retries, scale to markets per day, and add a cost cap.

9-minute read

Evaluate an LLM for 10-K extraction: build a labeled gold set, score field accuracy and faithfulness, test edge cases, weigh cost against quality.

9-minute read

Get reliable structured output from a finance LLM: define a strict schema, constrain generation, validate every field, and verify the numbers.

8-minute read

Select an LLM for a finance task by defining task, latency, cost, context, and quality constraints, then ranking on your own eval, not benchmarks.

8-minute read

Set up prompt caching for finance LLM workloads: identify the stable prefix, order the prompt for cache hits, manage the window, measure savings.

8-minute read

Summarize earnings calls with an LLM reliably: scope the summary, ground it in the transcript with citations, verify figures, estimate the cost.

8-minute read

Test an agent skill for a finance task: define the skill, run representative and adversarial inputs, score extraction, and measure cost and latency.

8-minute read