Factors | Docs | HyperPM

Write factors, test their predictive power with the four gates, and promote them for strategies to use.

Write a factor

First choose the axis:

Axis	What it is for	The question tested	Symbols
Cross-sectional	Selection: which coins to hold at a given moment	Do high-scoring symbols beat low-scoring ones	At least 10
Time-series	Timing: when to be in or out of one symbol	When the score is high, does the symbol rise more over the next few bars	One is enough

Cross-sectional

What it is for: Selection: which coins to hold at a given moment

The question tested: Do high-scoring symbols beat low-scoring ones

Symbols: At least 10

Time-series

What it is for: Timing: when to be in or out of one symbol

The question tested: When the score is high, does the symbol rise more over the next few bars

Symbols: One is enough

The axis cannot change once the factor is built. To try the other axis, click "Build along time" or "Build across instruments" on the factor page: the platform builds a new factor with the same definition; from then on the old and the new one are separate and do not affect each other.

Two ways to write one:

Say it in a conversation: for example "mine a BTC momentum factor"; the Agent proposes a factor, and once you confirm it builds it, runs the four gates and interprets the report.
Write it in the factor library: click "New factor" and fill in four steps.

Name it: a name and an English reference name (slug); strategies use it as factor://<slug>/v1.
Choose how to define it: a formula (fields and operators, such as MA(volume, 5) / MA(volume, 60), checked as you type; see the starter formulas on the right for reference); code (a factor class, for rules a formula cannot express; you can also click "Ask the agent" to hand it over); or a composite (two or more promoted factors on the same axis, weighted together).
Set the parameters: universe, bar size, horizon (how many bars ahead), range and so on all have defaults you can leave as they are. Preprocessing (trimming outliers, standardizing and so on) runs in the platform's fixed order.
Save: after "Save draft" it can still be edited; "Save and analyse" runs the four gates straight away.

New factor: name, axis and formula on the left, starter formulas on the right

Run the four gates

Each version is in one of three states:

Draft

A version not yet analyzed

Edit the definition, preprocessing and parameters; run the analysis

Analyse

Locked

A version that has been analyzed

Read the report, review, promote; the definition can no longer change

Promote

Promoted

A version that has been promoted

Used by strategies and composites; cannot change

Choose the version and the data on the factor page and click "Analyse" to run all four gates:

Gate	What it looks at
Output	How many symbols and bars are covered, how the scores are spread, the latest cross-section
Effectiveness	Strength and significance of the signal, stability (monthly and rolling IC), returns by group, whether it is tradable
Robustness	Whether it holds by year, by size (30-day turnover), by market regime and by universe; out of sample and on random subsamples
Relations	How closely it tracks the promoted factors in your library; time-series factors are also compared with the platform's known signals (momentum, reversal, volatility)

Output

How many symbols and bars are covered, how the scores are spread, the latest cross-section

Effectiveness

Strength and significance of the signal, stability (monthly and rolling IC), returns by group, whether it is tradable

Robustness

Whether it holds by year, by size (30-day turnover), by market regime and by universe; out of sample and on random subsamples

Relations

How closely it tracks the promoted factors in your library; time-series factors are also compared with the platform's known signals (momentum, reversal, volatility)

The top of a factor page: choose the version and the data, then "Analyse" ("Analyse again" on a version already run); the line above holds the four gates' verdicts

To change the definition, click "New draft" to start the next version from the current one; the old version and its report stay for comparison. If you only want a different horizon, bar size or universe, you can also name the factor in a conversation and say so.

Read the report

The row at the top of the report gives the four gates' verdicts, with a tab per gate:

A factor report: the four gates' verdicts in one row at the top, then a tab per gate with its flags and metrics

There are four verdicts: pass, warn, veto and not analysed. The effectiveness gate gives "veto" when the t value (|t|) is below the threshold: the signal is not significant enough and may be chance. The veto threshold starts at 2, and the platform raises it to suit the factor so that luck is not taken for a pattern. The report shows the threshold used.

The report's verdicts come from the platform's thresholds, not from the Agent. At the very bottom of each gate's report is a line of small text; open it to see each threshold's value and where it comes from.

Common flags:

Flag	What it means
IC weak / Rank ICIR weak	The prediction is weak, or comes and goes
IC win rate unstable	The direction is right on fewer than 55% of bars; it flips between right and wrong
turnover cost heavy	Scores change too fast and trading costs eat the return; try a longer horizon
low coverage	Many symbols get no score; check why the definition has so many gaps
small cross-section (fewer than 30 members)	The universe is too small; use a bigger one
survivorship bias in the universe	That stretch of history lacks contracts delisted later, so results may look better than they are
out-of-sample decay above 25%	The later part is clearly weaker
redundant with a known signal (momentum / reversal / volatility)	It is much like a known signal, so it adds little
redundant with a promoted factor	It is much like a factor you already promoted, so it adds little

IC weak / Rank ICIR weak

The prediction is weak, or comes and goes

IC win rate unstable

The direction is right on fewer than 55% of bars; it flips between right and wrong

turnover cost heavy

Scores change too fast and trading costs eat the return; try a longer horizon

low coverage

Many symbols get no score; check why the definition has so many gaps

small cross-section (fewer than 30 members)

The universe is too small; use a bigger one

survivorship bias in the universe

That stretch of history lacks contracts delisted later, so results may look better than they are

out-of-sample decay above 25%

The later part is clearly weaker

redundant with a known signal (momentum / reversal / volatility)

It is much like a known signal, so it adds little

redundant with a promoted factor

It is much like a factor you already promoted, so it adds little

The main figures in the report (cross-sectional):

Figure	How it is worked out, how to read it
Mean Rank IC	At each bar, the correlation between the ranking of the factor's scores and the ranking of later returns, averaged over all bars. The further from 0, the stronger the prediction
Rank ICIR	Mean ÷ standard deviation of Rank IC, for how steady the prediction is; when weak, the effectiveness gate warns
t	Significance of the mean Rank IC, allowing for overlapping holding periods
Hit rate	The share of bars where the predicted direction was right
Monotonicity	Whether group returns rise or fall in order of score; shown only, not part of the verdict
Long-short spread	Top group minus bottom group, average return over one horizon; not annualized, before costs
Turnover	Between two rebalances, the share of the top group that changes; the higher, the higher the trading cost
IC half-life	The number of bars at which predictive power halves as the horizon lengthens; how fast the signal fades
Out-of-sample decay	How much weaker the later part of the range is than the earlier part; not computed for ranges under a year

Mean Rank IC

At each bar, the correlation between the ranking of the factor's scores and the ranking of later returns, averaged over all bars. The further from 0, the stronger the prediction

Rank ICIR

Mean ÷ standard deviation of Rank IC, for how steady the prediction is; when weak, the effectiveness gate warns

Significance of the mean Rank IC, allowing for overlapping holding periods

Hit rate

The share of bars where the predicted direction was right

Monotonicity

Whether group returns rise or fall in order of score; shown only, not part of the verdict

Long-short spread

Top group minus bottom group, average return over one horizon; not annualized, before costs

Turnover

Between two rebalances, the share of the top group that changes; the higher, the higher the trading cost

IC half-life

The number of bars at which predictive power halves as the horizon lengthens; how fast the signal fades

Out-of-sample decay

How much weaker the later part of the range is than the earlier part; not computed for ranges under a year

Time-series reports use different figures, all along time: pooled TS-IC (each symbol's score against its own later returns, taken together), hit rate, breadth (the share of symbols pointing the same way), and the equity, Sharpe and drawdown of holding by the signal.

Review

In a review, the Agent questions the factor and looks for problems it may have. It reads the definition and the report (and the source text when there is one), only gives its assessment, changes nothing and does not affect promotion.

After at least the output and effectiveness gates have run, click "Review", or say "review <factor name>" in a conversation.

The result opens with an overall call: "Review: no objection", "Review: concerns", "Review: objection" or "Review: incomplete"; then what is good, the questions it cannot answer yet, and what to do before using the factor.

If the gates are run again, the earlier review is marked "review stale" and you can review again.

Promote

Click "Promote" (or say "promote <slug>" in a conversation) to make a version promoted.

All four gates must have run.
These need "I know, promote it anyway" ticked first: a veto from the effectiveness gate; a flag on it in the mining results table that needs acknowledging; a composite that does not clearly beat its best component; a time-series factor whose direction in the last stretch of the range is the reverse of the judged direction (unstable direction).

The panel that "Promote" opens: with an effectiveness veto, tick this box first, then "Promote" below can be clicked

Why a promoted version cannot change: strategies refer to one exact version, factor://<slug>/v<n>. If a promoted version could still change, the results of strategies using it could not be reproduced.

Strategies can only use promoted factors. If a request refers to a version that is not promoted yet, the platform tells you; reply "promote …" exactly as it shows to promote it. If your request named the factor by its reference name, it carries on by itself once promoted; otherwise send the request again.

Once promoted, the factor page shows "Build a selection strategy" ("Build a timing strategy" for time-series factors): it opens a conversation with a request to build a strategy from this factor already written; just send it.

Manage factors

Rename: changes only the display name; the reference name and strategies already using it are not affected.
Archive: an archived factor can only be viewed, renamed or have its description edited; restore it to make versions, analyze, review, promote or build strategies again. Factors cannot be deleted.
References: the "References" tab lists, for each version, the strategy versions and composites that use it. Factors mined from research documents also show their source.
Search: the list can be searched by name or slug, limited to promoted versions, and filtered by axis.