PoW
Proof-of-Work
SHA-256 nonce search where the network agrees the longest cumulative-work chain is canonical. Each block makes rewriting history exponentially harder.
while hash(block, nonce) > target:
nonce += 1
# avg: 10 min / block
A research playground for blockchain consensus,
zero-knowledge cryptography, and
quantitative trading systems.
We study the math that moves money at the speed of light.
From proof-of-work timestamping to BFT finality and zero-knowledge rollups — the cryptographic building blocks of trustless systems.
SHA-256 nonce search where the network agrees the longest cumulative-work chain is canonical. Each block makes rewriting history exponentially harder.
while hash(block, nonce) > target:
nonce += 1
# avg: 10 min / block
Validators stake capital as economic skin-in-the-game. Attestations + slashing replace hash power, punishing misbehavior with confiscated deposits.
reward = stake * uptime if equivocation: slash(stake, 1/3)
Tolerates up to f Byzantine nodes among n = 3f+1. Voting rounds guarantee safety even with a hostile leader, at the cost of message complexity.
quorum = 2f + 1 n >= 3f + 1 # safety
Prove a statement is true without revealing the witness. zk-SNARKs power private transactions and validity rollups that compress thousands of txs.
π = Prove(stmt, w) Verify(stmt, π) → ✓ // no w
Transactions reference multiple predecessors instead of a single block. Concurrent structure enables higher throughput at the cost of ordering rules.
tx.parents = [tips...] order := topological_sort(G)
Flood-fill propagation of transactions and blocks across the peer network. Each node validates before forwarding — trust minimized, locality maximized.
on_msg(m):
if valid(m) && !seen(m):
broadcast(m)
Satoshi's elegance: the chain with the most cumulative work is canonical — no coordinator, no voting, no central clock. An attacker rewriting n buried blocks must out-compute honest miners for that entire depth, and the probability of catching up drops exponentially. This is why six confirmations became the de-facto finality threshold.
P(attacker catches up from z blocks behind) ≤ (q/p)z
where q < p = 0.5 is the attacker's share of hash power. Below 50% the attack is a vanishing gambit; at or above it, the system collapses into a fork race.
Factor discovery, high-frequency microstructure, and reinforcement-learned execution — the machinery that turns market noise into Sharpe.
Cross-sectional signals — momentum, value, volatility, liquidity — neutralized against industry and style exposures to isolate pure alpha.
α = Σ wᵢ·fᵢ - β·mkt ret = rank(α), long top·k
Inventory-aware market making in the pro-rata LOB. Impulse + regular controls balance execution certainty against adverse selection risk.
dQ = λ(q)dt + σdW minimize: Var + penalty(inv)
Neural nets capture non-linear interactions traditional linear factors miss. Tensor-accelerated computation over 500–1000 engineered features.
ŷ = MLP(x₁..xₙ) loss = IC + L2·turnover
End-to-end trading agents learn policy from raw price/volume/news sentiment via Double DQN & GRPO — no handcrafted rules.
π* = argmax_π E[Σγᵗrₜ] s = [P, V, news_sentiment]
Cross-sectional neutralization + dynamic weight rebalancing. Markowitz core with risk parity overlay and turnover-constrained solvers.
min ½wᵀΣw - λμᵀw s.t. Σw = 1, |w|≤L
Volatility-targeted exposure, maximum-drawdown caps, and stress testing under regime switches keep capital survival the first invariant.
exposure = σ_target / σₜ if MDD > cap: de-risk
Crypto markets are the natural laboratory where these two worlds collide: a 24/7, fully-digital, API-native order book where on-chain settlement is the clearing layer. News-aware reinforcement agents now ingest LLM sentiment scores alongside tick-level price/volume, and the same DAG consensus research that secures the ledger also shapes latency-sensitive execution.
The frontier isn't one discipline — it's their intersection: verifiable off-chain compute for backtests, ZK-attested alpha, and on-chain settlement of strategies that traditionally lived in a black box.
A curated reading list spanning the seminal to the cutting edge. Click through to the original sources.
Satoshi Nakamoto
The genesis paper. Solves double-spending without a trusted third party via proof-of-work and a peer-to-peer timestamp server. Nine pages that launched an asset class.
Zhou Kai, Chen Fu, Lu Tianyuan, Cao Huaihu
Comprehensive survey of consensus evolution with focus on BR-DAG algorithms. Proposes an evaluation framework across throughput, latency, and fault tolerance.
Yuxuan Lu, Chang Liu, Lanju Kong, Xiangyu Niu
Tackles dominant-replica failure, transaction interruption, and high latency by automatically switching between leader-BFT and leaderless-BFT based on network conditions.
Md. Mainul Islam, Mpyana M. Merolec, Hoh Peter In
A PoA variant using verifiable random functions for fair, unpredictable leader selection. Outperforms Aura and Clique on throughput and consensus time across a six-node testbed.
Yijia Xiao, Edward Sun, Di Luo, Wei Wang
Simulates a trading firm with LLM agents as fundamental, sentiment, and technical analysts plus Bull/Bear debaters. Beats baselines on cumulative return, Sharpe, and max drawdown.
Zhizhuo Kou, Holam Yu, et al.
A three-stage risk-aware multi-agent framework: LLM-generated alpha factors, multimodal evaluation, and dynamic weight optimization. 53.17% cumulative return on SSE50 (Jan 2023–Jan 2024).
Yimin Du
PyTorch-accelerated tensor factor computation, geometric Brownian motion augmentation, and cross-sectional neutralization. 20% annualized, Sharpe > 2.0 on A-shares (2010–2024).
Yimin Du
Five-module system: cross-sectional prediction, opening-signal mixture models, dynamic position sizing, grid-optimized stop-loss, and multi-granularity timing. 15.2% annualized, MDD < 5%.
Qing-Yu Lan, Zhan-He Wang, et al.
Feeds LLM-derived news sentiment + raw price/volume into RNN/Transformer policy nets. DDQN & GRPO on crypto markets beat benchmarks without handcrafted features.
Konark Jain, Nick Firoozye, et al.
Models limit order book events with calibrated Hawkes kernels and per-event order sizes. Spread stays positive; non-parametric calibration on NASDAQ replicates concave market impact.
Rene Carmona, Kevin Webster
Redefines the "informed trader" for HFT and rigorously derives LOB clearing conditions. Two distinct frictions identified; ignoring them misstates wealth exchanged by ~50%.
Fabien Guilbaud, Huyên Pham
Impulse + regular controls for market making and best execution in pro-rata LOBs. Integro quasi-variational inequality characterizes the value function; overtrading risk analyzed.
A simulated research console. Type help to list commands.
Telemetry from the research node. (Static demo data — no backend required.)