Trimage towers and the Han River, seen from inside the complex

Quantitative research & trading · Trimage, Seongsu-dong, Seoul · 37.5442° N, 127.0450° E

We trade only what we failed to refute.

StreetSmart Research is a small quantitative trading firm working from the Han River in Seongsu. We declare every hypothesis before we see its result, test it against a benchmark that holds the same exposure, and keep the rejections in a ledger. Most ideas do not survive. We trade the ones that do.

39,519
formulaic factors screened in one nested search
73,800
barrier configurations per sweep, 537 symbols
12
automated checks before a result enters the ledger
3
independent locks before a live order can leave

Company

A research desk that happens to trade

We study short- and medium-horizon structure in crypto perpetual futures, leveraged ETFs and their swaps, and the cross-section of liquid altcoins. The work is organised like a laboratory: one problem framework, a registry of strategy families, and gates that every idea passes in the same order.

Model
Proprietary capital. Paper and shadow books run beside every live sleeve.
Markets
Binance and OKX perpetuals, Bybit, US leveraged ETFs, Korean flows as a covariate.
Horizon
Sub-second execution to multi-day cross-sectional holds.
Registry
30 strategy families graded A to N. Thirteen are recorded nulls.
Research

Hypotheses with a price on them

Each idea is stated with its minimum detectable effect and the trial count of its family before any data is touched.

Validation

Exposure-matched, out of sample

Purged cross-validation, placebo signals, and the expected maximum Sharpe of all trials run so far.

Execution

Costs measured, not assumed

Fee, half-spread and square-root impact, calibrated on our own fills to within half a basis point.

Risk

Shadow by default

Every strategy runs on a shadow book first. Reconciliation, alerts and a kill path stay on around the clock.

Problems we study

Six questions, and where each one stands

These are the problems we return to most often. The mathematics is the standard literature, restated in our notation. The verdicts are ours.

P1

Where does a Sharpe ratio come from?

We reduce every candidate to hit rate, magnitude, cost and breadth. If it cannot say which term it improves, it is not a strategy yet.

Where does a Sharpe ratio come from?
Framework  A 52.5% coin needs breadth. On our cross-sectional book, 27 nominal bets behave like about 20 independent ones.
P2

Is the best backtest just the luckiest?

Searching produces winners by construction. A discovery has to beat the maximum that noise alone would deliver over the same number of trials.

Is the best backtest just the luckiest?
Rejected 2  A self-improving search over 502 variants topped out at 0.61 against a null of 0.75. A 10,368-cell ladder grid reversed under conservative fills.
P3

Does the price revert, and how fast?

Mean reversion is only tradable if its half-life is shorter than a random walk’s would appear to be, after costs.

Does the price revert, and how fast?
Null  Optimal trading rules on intraday OU fits: positive in-sample, negative on hold-out. The fitted half-life was indistinguishable from shuffled returns.
P4

What will this trade actually cost?

Spread estimators built for daily equities measure volatility in crypto. Impact estimators built on signed flow measure cost.

What will this trade actually cost?
Holds  Against 196 of our fills: Hasbrouck ρ = +0.55, Amihud +0.42; Corwin–Schultz −0.01. Calibration κ = 0.178.
P5

Is it skill, or just more exposure?

Sizing rules, kill switches and regime gates change how much is held. A benchmark that holds the same average exposure removes that effect.

Is it skill, or just more exposure?
Mostly dials  Probability-based sizing tracked exposure at −0.98. Every capital schedule we rescored had ΔSR ≤ 0.
P6

Can a crash be seen coming after a run-up?

We forecast the probability of a five-day drawdown in the worst decile, from volatility alone and with every run-up signal we have.

Volatility only  Volatility alone beat the base rate on SOXL, TQQQ and BTC. Adding run-up history, or combining all signals into one model, made forecasts worse. Crashes followed falls and high volatility, not streaks of gains.

Method

Seven gates, in a fixed order

An idea moves to the next gate only with a written record of the last one. Nothing skips ahead, and a rejected design is never re-run on the same data to rescue it.

G0
Hypothesis
Mechanism, family, minimum detectable effect.
G1
Preregistration
Declaration hashed before results exist.
G2
Independent review
A second model challenges the design. Two holds force a redesign.
G3
Hold-out test
Exposure-matched benchmark, placebo, E[max SR].
G4
Paper
Real quotes, real fees and funding, shadow fills.
G5
System fit
Judged by its effect on the whole book, not alone.
G6
Live
Small, capped, armed by the CIO in person.

Principles

Three rules we do not bend

R1

Exposure-matched benchmarks

A strategy is compared with a passive position of the same average size. Lower drawdown from holding less is not an edge.

R2

The system is the unit

A component is adopted for what it does to the whole book: change in Sharpe, realised ratio, losses avoided, hypotheses blocked.

R3

Second-order evidence

Year-by-year stability, the equity curve and liquidity-conditioned IC are reported every time, even when the headline passes.

Technology & controls

Built so that the default is to do nothing

Our terminal separates the interface from a headless core, stores every quote and fill locally, and runs each algorithm on a shadow book until it is deliberately promoted.

01Ingest. Raw sources stored with their availability time, so no backtest can see the future.
02Research engine. Signal adapters, out-of-sample meta-models, placebo and no-information verdicts.
03Check registry. Twelve checkers gate the ledger. The first run caught a missing sweep record.
04Core. Market state, algorithms and shadow books in one headless process.
05Order management. Cost model, transaction-cost analysis and daily reconciliation that halts on mismatch.
06Monitoring. Hourly alerts, drift checks, nightly backups with restore tests.
LOCK 1

Exchange keys are stored with owner-only permissions and used only for signing. They are never displayed or transmitted.

LOCK 2

The live layer refuses to send until the CIO arms it, with per-order, per-position and daily-loss limits.

LOCK 3

An explicit live flag on the order path. A single command cancels everything and stops every algorithm.

Instruments

The terminal we actually trade from

Our own trading terminal: the screen and the engine are separate programs, and the overlays on the chart are the studies in our compendium, turned into instruments.

StreetSmart terminal
Tour the terminal

Publication · Vol. I

A Compendium of Results, Refuted and Otherwise

Results from mathematical finance, each restated as a definition or proposition, followed by a remark and its empirical status on our own data.

Contact

By introduction

We are a proprietary desk and do not take outside capital. We are glad to compare notes on methods, especially on results that failed for us and may work for you.

Read the research compendium
Desk
Trimage, Seongsu-dong, Seongdong-gu, Seoul
Hours
Markets never close. Neither does monitoring.
Archive
Earlier essays are being reorganised.

© 2026 StreetSmart Research. This page describes research methods and is not an offer, solicitation or investment advice. Past and simulated results do not indicate future returns. Live model performance is deliberately not disclosed. Skyline illustration drawn in code; not a photograph. Privacy