Engine running Session 1,284 UTC 00:00:00 Strategies 4 active Gross 0.82× Loss halt armed

Systematic trading research

Most patterns
are noise.
We look anyway.

Pandan Research searches a high-dimensional feature space for relationships that hold out of sample, then builds strategies from the few that survive. Four run in the book today. Several hundred did not make it out of the search.

Feature space · 3 of 41 dims Scanning
Pairs tested 000000 Retained 0000 Illustrative
Strategy contribution · 240 sessions Hover to inspect
Net of modeled costs

How a session runs

Four stages, in order, every time.

Each stage hands the next a single typed output. A stage cannot skip, retry a different way, or be talked into an exception. Every strategy walks the same path independently.

01 / Ingest

Data

Prices, volume, corporate actions and calendar events, normalized to one timeline. Bad ticks are quarantined rather than smoothed away.

in  →  vendor feeds
out →  bar table
02 / Extract

Signal

A short list of tested features produces one score per instrument, per strategy. Anything that survives only in-sample never reaches this stage.

in  →  bar table
out →  score vector
03 / Size

Risk

Scores are netted across strategies, then converted to positions through volatility targeting and hard exposure caps. The sizer can always return zero.

in  →  score vector
out →  target book
04 / Execute

Orders

Limit orders staged against liquidity, with slippage measured against arrival price and written back into the record before the next session opens.

in  →  target book
out →  fills + costs

Method

Built to be doubted.

The firm assumes its own edge is small, decaying, and possibly imaginary. Everything below follows from that assumption.

Out-of-sample or nothing

Every strategy is fit on one period and judged on another it has never seen. Walk-forward results are the only ones reported, including the bad ones.

Costs modeled first

Commission, spread, borrow and realistic slippage are subtracted before a strategy is evaluated — not added to a curve that already looks good.

Uncorrelated by construction

Strategies are admitted on the strength of what they add to the book, not their own curve. A profitable strategy that duplicates an existing one is rejected.

Default limits

What the engine will not do.

Limits are set before a session opens and enforced in the sizer, not in a policy document. Changing one requires a new deployment and leaves a record.

Maximum single position4.00% NAV
Maximum sector exposure20.00% NAV
Gross leverage ceiling1.00×
Daily loss halt1.50% NAV
Per-strategy risk budget25.00% of vol
Order types permittedLimit only
Discretionary overridesNone

Access

Currently in private testing.

Pandan Research is running paper and small live allocations while the execution layer is measured. Leave an address and you'll hear when the next group opens.

One email when access opens. Nothing else.