Tianrui Wu Resume PDF
Menu

Quant Lab / Experiment 01

Trading Systems and Execution

How can market-system research be made technically useful while separating observed evidence, published methods, validation targets, and counterfactual scenarios?

Updated July 15

Experiment 01 / 01

Deep Limit-Order-Book Architecture

How can spatial depth and temporal state be combined in a short-horizon LOB model?

The schematic follows the published DeepLOB design: a 100 × 40 depth tensor, convolutional feature extraction, a multi-scale Inception module, 64 LSTM units, and three-class softmax output.

10 LOB levelsCNN + InceptionLSTM 64
DeepLOB convolutional, Inception, and LSTM architecture One hundred limit-order-book snapshots with ten price levels and four price-volume fields pass through convolutional blocks, a multi-scale Inception module, 64 LSTM units, and a three-class softmax output. INPUT · 100 × 40 100 recent LOB states pᵃvᵃpᵇvᵇ L1L10 CNN STACK 1×2 @16 · stride 2 4×1 @16 ×2 depth collapse price / size → depth INCEPTION @32 1×1 → 3×1 conv 1×1 → 5×1 conv 3×1 pool → 1×1 concat multi-scale time features LSTM 64 temporal state SOFTMAX DOWN FLAT UP REFERENCE ARCHITECTURE · ORIGINAL SCHEMATIC BASED ON ZHANG, ZOHREN & ROBERTS (2018)
DeepLOB — Zhang, Zohren & Roberts (2018).

Experiment 01 / 02

Market Making and Backtesting

What did the sanitized 120-observation return path and drawdown look like?

120 daily observations compound to 83.2% with a maximum drawdown of -6.8%.

Observed data0% baseline7-day EMA
Market-making cumulative return The exact compounded path and its seven-day exponential moving average across 120 observations. FEB 2026MAY 2026
Bright line: 7-day EMA. Faint line: exact observed path. Download 120 observations.

Experiment 01 / 03

Optimal Execution

What do the stored run, TWAP comparison, and fee record actually establish?

Mean signed slippage was +3.22 bps versus the stored TWAP baseline and +2.34 bps above the matched reference. Aggregate fee divided by executed volume was -1.79 bps under the stored sign convention.

31 observed daysTWAP baseline2.34 bps gap-1.79 bps fee rate
Execution return, TWAP slippage, and observed fee rate The upper panel shows the exact 31-day compounded return and five-day exponential moving average. The lower panel shows observed mean signed slippage endpoints relative to a zero TWAP baseline and the separately observed aggregate fee rate. 31-DAY VWAP-RELATIVE RETURN · SHARED SCALE OCT 15, 2025NOV 14, 2025 OBSERVED AGGREGATE ENDPOINTS · BPS SELECTED +3.22 · VS TWAP MATCHED +0.88 FEE RATE -1.79 · SEPARATE TWAP 0.00 +2.34 GAP
Upper: exact daily relative-return path plus 5-day EMA. Lower: stored aggregate endpoints. The fee rate is observed, but fee reduction versus TWAP is unavailable because no TWAP fee series was retained. Inspect metrics.

Experiment 01 / 04

Microstructure Factor Research

Which common LOB factors are testable, and what should a strong validation profile look like?

The equations make the candidate signals inspectable. The IC/IR term structure is an explicitly illustrative acceptance target—not an observed or production performance series.

Queue imbalanceMicropriceOrder-flow imbalanceIllustrative IC / IR target
QUEUE IMBALANCEQIt = (qb1,t − qa1,t) / (qb1,t + qa1,t)
MICROPRICEμt = (pa1,tqb1,t + pb1,tqa1,t) / (qb1,t + qa1,t)
ORDER-FLOW IMBALANCEOFI[t,t+Δ] = Σn enen: signed best-queue change from order-book events
TRADE-FLOW IMBALANCETFIΔ = (QbuyΔ − QsellΔ) / (QbuyΔ + QsellΔ)

ILLUSTRATIVE ACCEPTANCE TARGET · NOT OBSERVED PERFORMANCE

Illustrative IC and information-ratio term structure A clearly labeled illustrative acceptance target shows information coefficient declining from 0.142 at one second to an assumed 0.0409 at five minutes and information ratio declining from 3.4 to an assumed 2.1. These values are not observed results. 0.16 0.0409 0 IC · SPEARMAN(Ft, Rt+h) 3.8 2.1 0 1s5s15s30s60s5m ICIR · MEAN(IC) / SD(IC)
IC 0.142 at 1s → assumed 0.0409 at 5m ICIR 3.40 at 1s → assumed 2.10 at 5m
Equations are common research definitions; OFI follows the best-quote event framing in Cont, Kukanov & Stoikov. The IC/IR paths are validation targets only and must be replaced by timestamped out-of-sample estimates before being described as results.

Experiment 01 / 05

Portfolio Risk

How much would a fivefold BTC exposure cut have changed the October 2025 crash path?

Using 1,441 public Coinbase five-minute closes, the full-exposure path reached a -11.9% maximum drawdown. A pre-event allocation of 20% BTC plus 80% cash reached -2.4% under the same price path.

Coinbase BTC-USD5-minute closes20% BTC + 80% cashCounterfactual
BTC October 2025 drawdown under full and twenty-percent exposure Coinbase BTC-USD five-minute closing prices from October 9 through October 14, 2025 are converted into drawdowns for a full BTC allocation and a counterfactual allocation of twenty percent BTC plus eighty percent cash. Maximum drawdowns are 11.9 percent and 2.4 percent respectively. 0%−4%−8%−12% OCT 9 2025OCT 14 2025
100% BTC maximum drawdown -11.9% 20% BTC + 80% CASH maximum drawdown -2.4% Close trough Oct 10, 21:15 UTC
E100t = Pt / P0 E20t = 0.80 + 0.20(Pt / P0) DDt = Et / maxs≤t(Es) − 1
Exact close-to-close paths from public Coinbase BTC-USD candles; 20% exposure is a no-trade allocation counterfactual set before the window. It excludes altcoins, derivatives, fees, funding, and intrabar marks. Download 1,441 candles.