How qex Evaluates¶
Every value is a table of dates by stocks¶
Every part of a qex expression stands for a panel: one value per date and per stock. close() is each stock's
close on each day; returns(5) is each stock's five-day return on each day; rank(returns(5)) is each stock's rank
among that day's stocks. A number such as 0.5 is the same value on every date and stock.
So an expression is not a calculation for one stock or one day. It is one formula applied to the whole panel at once, and its result is again a panel: a signal, ready to become a book.
An expression is a tree of operators, read from the inside out:
rank
└── ÷
├── − returns(5) each stock's 5-day return, negated
└── volatility(20) each stock's 20-day volatility
A part used twice, by name with let or written out twice, is the same part: qex recognizes it and computes it once.
Four kinds of operator¶
| Kind | Works | Examples | A stock's result depends on |
|---|---|---|---|
| Element-wise | on each date and stock separately | +, log, abs, where, comparisons |
its own values that day |
| Time-series | down each stock's history | ts_mean, delay, ts_rank, ewma, returns |
its own past |
| Cross-sectional | across the stocks on each day | rank, zscore, demean, regression_neut |
the other stocks that day |
| Group | across the stocks in its group on each day | group_demean, group_rank |
the stocks sharing its label |
Most signals mix all of them. In rank(-returns(5) / volatility(20)), returns and volatility look down each
stock's history, / and - work element by element, and rank compares the stocks on each day.
Two kinds of regression¶
The direction matters most for regressions, because qex has both kinds, and they answer different questions.
| Time-series regression | Cross-sectional regression | |
|---|---|---|
| Operators | ts_beta, ts_beta_shrunk, ts_alpha, ts_resid and the other ts_ pair operators |
regression_neut, regression_alpha, regression_beta |
| One fit for | each stock, on each day | each day |
| Its data points are | the stock's last period days |
the day's stocks |
| The question | how does this stock move with x, through time? |
across today's stocks, how much of y lines up with x? |
x is |
one or more series over time, such as the market's return or factor returns | one or more values for each stock that day, such as each stock's beta |
| The coefficients | belong to the stock: each stock has its own beta and alpha | belong to the day: one intercept and one slope, shared by every stock |
They are often used together. A time-series regression estimates each stock's beta from its history; a cross-sectional regression then takes those betas out of today's signal:
let beta = ts_beta_shrunk(returns(1), returns(1, 'market'), 252) # through time
in regression_neut(rank(-returns(5)), beta) # across stocks
Their residuals differ in the same way. ts_resid(returns(1), returns(1, 'market'), 252) is the part of a stock's
return today that the market's move doesn't explain, judged against that stock's own history.
regression_neut(y, beta) is the part of today's y that beta doesn't explain, judged across today's stocks.
Which stocks a cross-section compares¶
A cross-sectional or group operator compares the stocks in the universe that day, and gives every other stock no value. Time-series operators are not restricted: a stock that joined the universe today still has its past returns, volatility and beta. See Universes.
One consequence: a time-series operator applied to a cross-sectional result, such as ts_mean(rank(x), 20), has no
history for a stock from before it joined. Smooth first, then compare, where the meaning allows:
rank(ts_mean(x, 20)) has a value on a stock's first day in the universe.
Windows and history¶
A time-series operator looks back over a window, and needs enough values in it. qex's operators follow three rules:
| Rule | Operators | Gives a value once | On a day its input is missing | After a gap |
|---|---|---|---|---|
| Part of the window | ts_mean, ts_std, ts_sum, ts_max, ts_min, ts_var, ts_quantile, ts_corr, ts_cov, ts_beta, ts_beta_shrunk, volatility, volume, adv |
the window has min(20, period) values, or the min_periods given to an operator that takes one |
still gives a value, from the rest of the window | recovers at once |
| The whole window | ts_rank, ts_skew, ts_kurt, ts_argmax, ts_argmin, decay_linear |
every day of the window has a value | no value | no value until the gap has left the window |
| Exponential | ewma, ewma_var, ewma_cov, ts_beta_ewma |
its first value | still gives a value | carries straight on |
So ts_mean(x, 252) has a value 20 days after a listing, from 20 days; ts_rank(x, 252) waits a full year, and one
missing day then hides it for a year. When a window's minimum matters, ask for it explicitly: ts_corr, ts_cov,
ts_beta, ts_beta_shrunk and ts_quantile take it as an argument, and
where(ts_count_events(x, 252) >= 200, signal, None) works for any operator.
ts_count_events(x, period) counts the values x has in its window, so it has an answer on every day once period
days have passed. The other event operators (ts_mean_events, ts_std_events, ts_sum_events) look back over the
last n values that exist, however far back, for data such as quarterly earnings. See the
operator reference.
Point in time¶
A value on day t uses data up to and including day t, never later. There is no operator that looks forward: delay
and delta take only non-negative periods. A book trades on the signal with a delay, normally of one day, so a
signal computed from day t's close is first traded at day t+1's close.
Two things can still let the future in, and qex checks for both:
- Adjusted price levels fold in later splits and dividends. Comparing them across stocks, or with a fixed number, sees the future. See Adjusted Data and Look-Ahead.
- Labels: a stock's group on a past day must be the label it had then. Labels given on fewer dates than the prices hold from their date forward, never backward. See Groups and Labels.
Long and short¶
A signal's values become a book only after the book demeans and scales them (see the Python section). What the signal says is therefore relative: a stock is held long or short according to where its value sits against the day's other values.
rank(x) gives values in (0, 1], so rank(x) and rank(x) - 0.5 make the same book. A 0 is a value like any
other, so a stock at 0 is held against the day's average; a missing value is left out of the book. That difference
runs through the whole language: see Missing Values.
Equal long and short dollars don't make a book market-neutral; see Market Beta.