LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
P and Q name two jobs in quantitative finance, as laid out in a short paper the lesson reviews, "P vs. Q," associated with Antonio Meucci. Q work extrapolates prices that already exist. P work tries to say something statistically useful about the future. The lesson treats the split as a clear teaching table and a blurry career. Most people touch both. The table is still the fastest way to see what each side optimizes.
Why it matters
A model can be asked two different questions. One question is what an illiquid contract should be worth so that it lines up with liquid contracts today. The other is whether a forecast has a statistical case that could support trading over time. Mixing those questions produces the wrong test. A calibrated volatility surface is not a forecast, and a forecast error is not a calibration residual. The paper's table keeps the tests apart. The lesson's favorite overlap is a single options trade that uses a Q price as the signal for a P decision.
The idea in plain words
On the Q side the goal is to extrapolate the present. The probability is the risk-neutral measure Q. The processes are continuous-time martingales. The dimension is low: a handful of parameters, not hundreds of thousands. Working tools are Ito calculus together with partial differential equations, and the hard problem is calibration. The business is the sell side.
Listed prices are the raw material. A share quote and a one-month call struck near that share are already on the screen. An exotic with no quote is the thing to price. Whatever model is chosen, a Heston stochastic-volatility model, a jump model, Merton, Bates, or the SABR model the lesson flags as a popular stochastic-volatility choice, the parameters are fit so vanilla prices and the implied-volatility surface are matched as well as that model allows. That fit is the calibration. The output is a price meant to be consistent with the liquid market right now.
The P side is buy-side work under real-world probabilities and discrete series. A cash book can span 500 to 1,000 names. Multivariate statistics are the toolkit, and evaluation is the hard problem. The stated aim is a forecast with a statistical case for lasting profitability. The table names that aim.
Frictions live here. Transaction costs, gaps, and liquidity are often dropped in a continuous Q model. A P book cannot drop them. The lesson's illustration is an illiquid name that rises 200 percent on paper while no buyer exists. The gain cannot be realized. Theory that ignores the buyer has answered a different problem.
How to use it
Use the table as a checklist before you pick a tool. If the deliverable is a price for something illiquid, start from quoted vanillas, pick a low-dimensional continuous model, and judge yourself by calibration error. If the deliverable is a trading forecast, start from discrete panels, expect a wide cross-section, and judge yourself by an evaluation that survives costs.
The overlap example is statistical arbitrage in the lesson's simplest picture. A Q model prices a listed option below the price on the screen. The trading decision treats that gap as a signal that the quote is rich relative to the model, sells the option, and looks to close if the quote moves to the model value. The lesson calls the gap alpha only inside that story. It immediately lists what the story leaves out: transaction costs, liquidity, and whether the gap still appears in an out-of-sample backtest. Those caveats are part of the example.
The paper also compares continuous and discrete representations of ideas such as volatility clustering, including an autoregressive moving-average process against a stochastic differential equation, and Lévy or jump processes against Brownian motion. The lesson notes that material and spends its attention on the arbitrage picture instead.
What usually goes wrong
A hard line between the jobs is the mistake the lesson corrects. The table is a map, and a working desk crosses it. A calibration is not a forecast, and a backtest is not an arbitrage-free exotic price.
The 200 percent illustration fails if no buyer exists. A model residual smaller than the cost of hedging it is not yet a trade. A Q model with a huge parameter count, or a P model with only a handful of parameters on a thousand-name panel, fights the sizes the table describes.
Questions people ask
What does the Q table optimize?
Calibration to current liquid prices, with low-dimensional continuous models, Ito calculus, and partial differential equations, on a sell-side desk.
What does the P table optimize?
Evaluation of forecasts on large discrete panels, with multivariate statistics, where costs and liquidity count, on a buy-side desk.
Where do the two meet in the lesson?
When a model price and a live option quote disagree, and a trading book treats the gap as a signal, subject to costs, liquidity, and an out-of-sample test.