A self-directed quantitative finance research project
I am building this lab to apply my software engineering, machine learning, statistics, and data systems background to questions in quantitative finance. Its strategy lab tests simple rules against historical prices and asks what the evidence can—and cannot—support. A separate options calculator examines theoretical prices under stated assumptions.
A short field guide · No finance background needed
Begin with a testable idea. Follow the data, inspect the assumptions, and keep the limitations in view.
Research process, not a performance chart.A backtest asks: “What would have happened if I had followed this rule in the past?”
You choose a period, a starting amount of imaginary money, and a rule. The lab walks through historical prices in order, simulates buying or selling, and tracks the portfolio’s value.
Quantitative research simply means using numbers and repeatable rules to investigate an idea. Here, you can inspect the decisions, compare the outcome, and keep a record of what you tested.
My background is primarily in computer science and software engineering. I built Orion to develop the mathematical, statistical, and financial research skills that quantitative finance requires, through small experiments I can explain and repeat.
The lab starts with SPY, an exchange-traded fund: a basket of investments that trades like a stock. SPY tracks the S&P 500, an index of large U.S. companies.
Buy at the first selected closing price. Keep the investment until the experiment ends.
“What if I simply stayed invested?”
This is the benchmark—the simple alternative your idea is compared against. It receives the same starting capital, dates, asset, and cost assumptions.
Own the asset when its previous closing price is higher than its close a chosen number of sessions earlier. Otherwise, hold cash.
“Does a recent rise tend to continue?”
The lookback is that number of trading sessions. A 30-day lookback means 30 market observations, not 30 calendar days. This rule can miss recoveries or trade repeatedly in a choppy market.
The moving-average filter owns the asset when yesterday’s close is strictly above the average of the previous N closes, including yesterday. Otherwise it holds cash. The decision uses only earlier prices and takes effect at today’s close.
“Is the latest close above its recent trend?”
This gives a second, testable definition of a trend. It shares the same asset, benchmark, transaction-cost setting, and limitations as the momentum rule.
You can also import your own daily CSV. The benchmark then holds that same imported asset, rather than SPY.
The engine must only use information that was already available. Seeing tomorrow’s price today would make the experiment unfair.
Momentum compares yesterday’s close with one lookback earlier; the moving-average filter compares it with a recent average. Earlier data warms up either rule before your selected test window.
If the rule changes from cash to invested, buy. If it changes back, sell. Deduct the selected cost on each transaction.
A purchase at today’s close first earns the return from today to tomorrow. Each day’s cash plus investment value becomes a point on the chart.
Suppose Monday closes at $100 and Tuesday at $110. A one-session momentum rule sees that rise and buys at Wednesday’s close. It cannot claim the gain from Monday to Tuesday—or any movement earlier on Wednesday—because it did not own the asset then.
Daily SPY prices come from Yahoo Finance. A weekday refresh is scheduled for 23:30 UTC: 7:30 p.m. Eastern in summer and 6:30 p.m. in winter. Today’s prices are excluded until 6 p.m. Eastern. The lab displays the latest available date and warns when data is old.
Prices are adjusted for dividends and stock splits so comparisons use a consistent basis. They are research values, not the exact historical prices at which you could have traded.
You can own fractional units or hold cash. Cash earns no interest. The model has no borrowing, taxes, or separate estimate for the difference between expected and actual trade prices. Open holdings are valued at the final close without a forced sale.
Calculations run in your browser. The data source is unofficial; a failed refresh keeps the last published dataset available. “Check for data update” checks that published snapshot, not live quotes.
The portfolio chart shows how your imaginary money changed over time. The solid line follows your rule; the dashed line follows buy and hold.
Starting with $100 and ending with $120 means a 20% return. The lab includes modeled trading costs.
If the portfolio reaches $120, then falls to $90, that decline is 25%. A later recovery does not erase the experience of that fall.
This measures the variation in daily returns, scaled to a year. It counts both upward and downward movement; it is not a worst-case loss.
This compares average daily return with daily volatility, scaled to a year. Here the cash reference rate is zero. Higher values indicate more return per unit of variation in this sample, not a guarantee of a better strategy.
Total return = final value ÷ starting capital − 1. Volatility = sample standard deviation of daily returns × √252. Sharpe = mean daily return ÷ that daily standard deviation × √252. The lab assumes 252 trading sessions per year and reports Sharpe as N/A when variation is zero. Drawdown compares each value with the highest value reached so far, including starting capital.
Try this: “How does a 30-session momentum rule compare with simply holding SPY?”
Use the default dataset and “Time-series momentum.” Start with a lookback of 30. For your first run, the default dates provide enough earlier history to warm up the rule.
Start with $100,000 of simulated capital and 5 basis points per trade. A basis point is 0.01%, so 5 bps means a $5 cost for every $10,000 bought or sold.
Click “Run backtest.” Compare return and drawdown with the benchmark. Then inspect Trade history to see when the rule bought, sold, or stayed in cash.
Try a 60-session lookback, keeping everything else the same. Did trading become less frequent? Did the portfolio miss a recovery? Changing one setting makes the comparison easier to understand.
“Save experiment” keeps a run in this browser on this device, up to the latest 10. “Export experiment JSON” downloads the data, settings, and results. “Export equity CSV” downloads daily values you can open in a spreadsheet. Clearing browser storage removes local saves.
Choose “Import daily CSV,” enter an asset label, and select a file with
date,close columns. Use dates like 2026-01-02, positive
decimal prices, and one row per daily observation, oldest first. Dates must be
unique. Use a consistent adjustment basis and include enough observations before
your start date for the momentum lookback.
The lab accepts 3–20,000 rows in a file smaller than 2 MB. It validates the format but cannot independently verify your provider’s prices or trading calendar.
Orion is a working learning project, not a record of live trading. Its current scope is deliberately narrow so that each calculation and assumption can be inspected.
A static HTML/CSS and browser JavaScript interface runs separate JavaScript backtest and options-pricing engines. Python scripts validate and refresh the published daily dataset. GitHub Actions checks the lab; the portfolio's scheduled workflow refreshes the published data. The portfolio serves the reviewed lab build from a Dockerized Nginx site.
The current code does not use a server API, database, React, or a machine-learning model. Those would be separate future design decisions, not existing capabilities.
THE NEXT EXHIBIT IS YOURS
The lab is a place to learn how an idea behaves—and where it breaks.
Enter the research labExplore the European options pricer ↗
Historical backtests. Theoretical option values. Assumptions you can inspect.