<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>Marketmaker.cc Blog</title>
  <link>https://marketmaker.cc/zh-Hant/blog/1/</link>
  <atom:link href="https://marketmaker.cc/zh-Hant/blog/rss.xml" rel="self" type="application/rss+xml" />
  <description>AI trading, market microstructure, and DeFi insights from Marketmaker.cc.</description>
  <language>zh-Hant</language>
  <lastBuildDate>Thu, 20 Aug 2026 00:00:00 GMT</lastBuildDate>
  <item>
    <title>Scoring Probabilistic Forecasts: CRPS, PIT Calibration, and DeepAR</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/probabilistic-forecasting-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/probabilistic-forecasting-trading/</guid>
    <pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate>
    <description>How to evaluate a predictive distribution honestly — CRPS as a proper scoring rule, the PIT histogram as a calibration diagnostic, and DeepAR sampling in GluonTS.</description>
    <category>forecasting</category>
    <category>probabilistic</category>
    <category>quantile-regression</category>
    <category>DeepAR</category>
    <category>uncertainty</category>
    <enclosure url="https://marketmaker.cc/images/blog/probabilistic-forecasting-trading-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Physics-Informed Neural Networks for Options Pricing</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/pinn-options-pricing/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/pinn-options-pricing/</guid>
    <pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate>
    <description>Putting the Heston, free-boundary, and jump-diffusion pricing PDEs into a neural network loss — the log-price residual, the mixed-partial autograd trick, and what still needs measuring.</description>
    <category>deep-learning</category>
    <category>PINN</category>
    <category>options</category>
    <category>Black-Scholes</category>
    <category>PDE</category>
    <enclosure url="https://marketmaker.cc/images/blog/pinn-options-pricing-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>PCMCI: Causal Discovery in Multivariate Crypto Time Series</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/pcmci-causal-discovery-crypto/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/pcmci-causal-discovery-crypto/</guid>
    <pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate>
    <description>How PCMCI&apos;s two-stage MCI test recovers directed causal links between crypto assets where correlation and bivariate Granger cannot — the construction, the tigramite pipeline, and the real-data study it still needs.</description>
    <category>causal-inference</category>
    <category>PCMCI</category>
    <category>causal-discovery</category>
    <category>time-series</category>
    <category>crypto</category>
    <enclosure url="https://marketmaker.cc/images/blog/pcmci-causal-discovery-crypto-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Order Flow Imbalance: The Cont-Kukanov-Stoikov Event Decomposition</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/order-flow-imbalance-prediction/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/order-flow-imbalance-prediction/</guid>
    <pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate>
    <description>Turning raw book updates into a signed flow quantity: the CKS event decomposition, multi-level OFI with PCA reduction, and Lee-Ready trade classification — plus an honest accounting of what the headline R-squared actually measures.</description>
    <category>microstructure</category>
    <category>order-flow</category>
    <category>imbalance</category>
    <category>prediction</category>
    <category>HFT</category>
    <enclosure url="https://marketmaker.cc/images/blog/order-flow-imbalance-prediction-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Neural ODEs: Does Continuous Time Beat a Delta-Time Feature?</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/neural-ode-continuous-finance/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/neural-ode-continuous-finance/</guid>
    <pubDate>Sun, 16 Aug 2026 00:00:00 GMT</pubDate>
    <description>Neural ODEs, Neural SDEs and continuous normalizing flows for irregularly-sampled market data — and the one ablation that decides whether continuous dynamics are worth their solver cost.</description>
    <category>deep-learning</category>
    <category>neural-ODE</category>
    <category>continuous-time</category>
    <category>SDE</category>
    <category>dynamics</category>
    <enclosure url="https://marketmaker.cc/images/blog/neural-ode-continuous-finance-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Multi-Task Learning for Simultaneous Price, Volume, and Volatility Prediction</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/multi-task-learning-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/multi-task-learning-trading/</guid>
    <pubDate>Sat, 15 Aug 2026 00:00:00 GMT</pubDate>
    <description>Does jointly predicting return, volume, and volatility actually help? Measuring loss-balancing schemes and diagnosing negative transfer through gradient cosine similarity — with a classical baseline and purged walk-forward folds.</description>
    <category>deep-learning</category>
    <category>multi-task</category>
    <category>shared-representation</category>
    <category>prediction</category>
    <category>auxiliary</category>
    <enclosure url="https://marketmaker.cc/images/blog/multi-task-learning-trading-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Model Pruning for Low-Latency Trading Inference</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/model-pruning-low-latency-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/model-pruning-low-latency-trading/</guid>
    <pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate>
    <description>Magnitude and structured pruning, the Lottery Ticket Hypothesis, movement pruning, distillation and 2:4 sparsity — the methods behind shrinking a trading model, and what still has to be measured before any of it ships.</description>
    <category>model-compression</category>
    <category>pruning</category>
    <category>lottery-ticket</category>
    <category>latency</category>
    <category>deployment</category>
    <enclosure url="https://marketmaker.cc/images/blog/model-pruning-low-latency-trading-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Updating the Volume Curve Intraday: Does Adaptive Forecasting Actually Help?</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/liquidity-prediction-execution/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/liquidity-prediction-execution/</guid>
    <pubDate>Thu, 13 Aug 2026 00:00:00 GMT</pubDate>
    <description>Our VWAP article shipped a static, weekly-refit volume curve and called the forecaster the weakest link. This is the follow-up: a Bayesian intraday updater run against the same 500-parent BTCUSDT harness, with the IS delta conditioned on realized curve error.</description>
    <category>microstructure</category>
    <category>liquidity</category>
    <category>execution</category>
    <category>prediction</category>
    <category>order-book</category>
    <enclosure url="https://marketmaker.cc/images/blog/liquidity-prediction-execution-hero.webp" type="image/webp" />
  </item>
  <item>
    <title>Koopman Operators and DMD: Do Market Modes Survive Out-of-Sample?</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/koopman-operator-market-dynamics/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/koopman-operator-market-dynamics/</guid>
    <pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate>
    <description>Dynamic Mode Decomposition fits a linear operator to nonlinear market dynamics. The only question that matters: do the fitted modes persist from one window to the next, and does the rolling spectral radius lead realised volatility? Here is the measurement protocol and the code to run it.</description>
    <category>mathematics</category>
    <category>Koopman</category>
    <category>dynamical-systems</category>
    <category>spectral</category>
    <category>prediction</category>
    <enclosure url="https://marketmaker.cc/images/blog/koopman-operator-market-dynamics.webp" type="image/webp" />
  </item>
  <item>
    <title>Knowledge Distillation: Compressing Trading Models for Low-Latency Deployment</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/knowledge-distillation-trading-models/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/knowledge-distillation-trading-models/</guid>
    <pubDate>Tue, 11 Aug 2026 00:00:00 GMT</pubDate>
    <description>The blog&apos;s standing answer to the accuracy-vs-latency tension is a two-stage fast/slow split. Distillation is a different answer: train one small model to mimic the ensemble. The KD loss, temperature, born-again nets, early exits for a variable latency budget, and the distill-to-FPGA pipeline — plus the measurements that would decide whether it beats the two-stage split.</description>
    <category>model-compression</category>
    <category>distillation</category>
    <category>latency</category>
    <category>HFT</category>
    <category>deployment</category>
    <enclosure url="https://marketmaker.cc/images/blog/knowledge-distillation-trading-models.webp" type="image/webp" />
  </item>
  <item>
    <title>The Other Way a Regression Lies: Endogeneity, 2SLS, and the Gamma Calibration Problem</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/instrumental-variables-quantitative-finance/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/instrumental-variables-quantitative-finance/</guid>
    <pubDate>Mon, 10 Aug 2026 00:00:00 GMT</pubDate>
    <description>Selection bias in the search is not the only way a regression fools you. When the regressor is correlated with the error, more data makes the estimate more confidently wrong. Instrumental variables applied to the one endogeneity problem this blog has already left open: permanent impact from net taker flow.</description>
    <category>causal-inference</category>
    <category>instrumental-variables</category>
    <category>2SLS</category>
    <category>econometrics</category>
    <category>endogeneity</category>
    <enclosure url="https://marketmaker.cc/images/blog/instrumental-variables-quantitative-finance.webp" type="image/webp" />
  </item>
  <item>
    <title>Hawkes Processes for Order Arrival and Market Event Modeling</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/hawkes-process-order-arrival/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/hawkes-process-order-arrival/</guid>
    <pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate>
    <description>Fitting a self-exciting point process to real crypto trade tape: where the three numbers (mu, alpha, beta) come from, how to estimate the branching ratio n, whether the exponential kernel survives a goodness-of-fit test, and how much n moves when you change the estimation window.</description>
    <category>microstructure</category>
    <category>Hawkes-process</category>
    <category>point-process</category>
    <category>order-arrival</category>
    <category>crypto</category>
    <category>HFT</category>
    <enclosure url="https://marketmaker.cc/images/blog/hawkes-process-order-arrival.webp" type="image/webp" />
  </item>
  <item>
    <title>Hamiltonian Neural Networks: Does a Financial System Conserve Anything?</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/hamiltonian-neural-networks-finance/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/hamiltonian-neural-networks-finance/</guid>
    <pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate>
    <description>Hamiltonian Neural Networks are provably stable — but stability is worthless if the conserved quantity does not exist. Learning a scalar H via autograd, symplectic integration, and the falsification test that decides whether a financial (q, p) pair is canonical at all.</description>
    <category>deep-learning</category>
    <category>Hamiltonian</category>
    <category>physics-informed</category>
    <category>dynamics</category>
    <category>conservation</category>
    <enclosure url="https://marketmaker.cc/images/blog/hamiltonian-neural-networks-finance.webp" type="image/webp" />
  </item>
  <item>
    <title>Toda-Yamamoto vs Differenced Granger: Does the BTC Lead-Lag Survive?</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/granger-causality-lead-lag-crypto/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/granger-causality-lead-lag-crypto/</guid>
    <pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate>
    <description>Granger causality on crypto prices done two ways — differenced returns and Toda-Yamamoto on levels — with a correct Wald implementation, an effective-N corrected causality matrix, and a rolling-stability test of whether the lag is tradeable at all.</description>
    <category>causal-inference</category>
    <category>Granger-causality</category>
    <category>lead-lag</category>
    <category>crypto</category>
    <category>VAR</category>
    <enclosure url="https://marketmaker.cc/images/blog/granger-causality-lead-lag-crypto.webp" type="image/webp" />
  </item>
  <item>
    <title>非參數價格建模的高斯過程</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/gaussian-processes-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/gaussian-processes-trading/</guid>
    <pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate>
    <description>金融時間序列的核設計——母體粗糙度、局部週期成分、光譜混合——以及作為不需要驗證集的正則化器的邊際似然。加上在任何東西可以交易之前仍然需要衡量的誠實清單。</description>
    <category>bayesian</category>
    <category>gaussian-process</category>
    <category>kernel</category>
    <category>uncertainty</category>
    <category>non-parametric</category>
    <enclosure url="https://marketmaker.cc/images/blog/gaussian-processes-trading.webp" type="image/webp" />
  </item>
  <item>
    <title>基於偏微分方程的金融建模的傅立葉神經算子</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/fourier-neural-operator-finance/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/fourier-neural-operator-finance/</guid>
    <pubDate>Wed, 05 Aug 2026 00:00:00 GMT</pubDate>
    <description>算子學習映射整個函數空間，而不是點——傅立葉神經算子如何在頻率空間中參數化 PDE 解算子，為選擇權定價購買什麼，以及它的哪些主張仍然需要在真實硬體上測量。</description>
    <category>deep-learning</category>
    <category>FNO</category>
    <category>PDE</category>
    <category>operator-learning</category>
    <category>options</category>
    <enclosure url="https://marketmaker.cc/images/blog/fourier-neural-operator-finance.webp" type="image/webp" />
  </item>
  <item>
    <title>一整天的內容能勝過十分鐘的內容嗎？ Flash 注意力與序列長度問題</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/flash-attention-trading-models/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/flash-attention-trading-models/</guid>
    <pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate>
    <description>Flash Attention 讓 23,400 步驟的交易上下文無需計算——平鋪、線上 softmax 和 N^2 d^2 / M 的 IO 界限。更長的上下文是否使模型更好是一個單獨的問題，它必須被測量，而不是斷言。</description>
    <category>deep-learning</category>
    <category>attention</category>
    <category>Flash-Attention</category>
    <category>GPU</category>
    <category>optimization</category>
    <enclosure url="https://marketmaker.cc/images/blog/flash-attention-trading-models.webp" type="image/webp" />
  </item>
  <item>
    <title>整合方法：結合弱學習者以獲得穩健的 Alpha</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/ensemble-methods-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/ensemble-methods-trading/</guid>
    <pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate>
    <description>本部落格未涵蓋有關交易整合的三件事：為什麼協方差項（而不是模型計數）在整合誤差中占主導地位，堆疊如何作為具有折疊元特徵的訊號組合層工作，以及跨多個訊號的周轉淨額是否真正降低了執行成本。</description>
    <category>machine-learning</category>
    <category>ensemble</category>
    <category>bagging</category>
    <category>stacking</category>
    <category>alpha</category>
    <enclosure url="https://marketmaker.cc/images/blog/ensemble-methods-trading.webp" type="image/webp" />
  </item>
  <item>
    <title>Double Machine Learning：估計因果參數而非預測報酬</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/double-ml-causal-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/double-ml-causal-trading/</guid>
    <pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate>
    <description>本博客至今所有模型回答「什麼預測什麼」。Double ML 回答「什麼導致什麼」——帶有可辯護的標準誤。部分線性模型、Neyman 正交性、order book 數據上的清理交叉擬合，以及為何有效 DML 信賴區間僅對一個預先指定的問題成立的誠實說明。</description>
    <category>causal-inference</category>
    <category>double-ML</category>
    <category>treatment-effect</category>
    <category>econometrics</category>
    <category>quant</category>
    <enclosure url="https://marketmaker.cc/images/blog/double-ml-causal-trading.webp" type="image/webp" />
  </item>
  <item>
    <title>交易中異質治療效果的因果森林</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/causal-forests-treatment-effects-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/causal-forests-treatment-effects-trading/</guid>
    <pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate>
    <description>本博客每個回測估計條件平均值。因果森林估計條件治療效果 — tau(x) 而非 mu(x) — 具有誠實分裂、自適應核權重表示以及告訴您發現的異質性是否真實的校準測試。</description>
    <category>causal-inference</category>
    <category>causal-forest</category>
    <category>heterogeneous-effects</category>
    <category>econometrics</category>
    <category>treatment</category>
    <enclosure url="https://marketmaker.cc/images/blog/causal-forests-treatment-effects-trading.webp" type="image/webp" />
  </item>
  <item>
    <title>認識與對立：衡量收益模型不知道的東西</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/bayesian-neural-networks-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/bayesian-neural-networks-trading/</guid>
    <pubDate>Fri, 31 Jul 2026 00:00:00 GMT</pubDate>
    <description>本博客的每個位置 sizing 規則都把不確定性坍縮成單一純量。MC Dropout 和深度集成將其拆分為模型無知與市場噪聲——這兩者值得不同的 position size。</description>
    <category>bayesian</category>
    <category>uncertainty</category>
    <category>neural-network</category>
    <category>risk</category>
    <category>position-sizing</category>
    <enclosure url="https://marketmaker.cc/images/blog/bayesian-neural-networks-trading.webp" type="image/webp" />
  </item>
  <item>
    <title>用於系統化交易流程的 AutoML</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/automl-systematic-trading/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/automl-systematic-trading/</guid>
    <pubDate>Thu, 30 Jul 2026 00:00:00 GMT</pubDate>
    <description>自動特徵生成（tsfresh、Featuretools）、具預算感知的模型搜尋（FLAML 的節省成本最佳化器），以及 WorldQuant 公式化 alpha 工廠——本部落格搜尋與過度擬合系列尚未涵蓋的研究流程部分，以及該系列自身結果對它們的啟示。</description>
    <category>AutoML</category>
    <category>NAS</category>
    <category>feature-engineering</category>
    <category>automation</category>
    <category>quant</category>
    <enclosure url="https://marketmaker.cc/images/blog/automl-systematic-trading/hero.webp" type="image/webp" />
  </item>
  <item>
    <title>將清算級聯作為交易訊號：解讀被迫且預先宣告的流量</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/liquidation-cascades-trading-signal/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/liquidation-cascades-trading-signal/</guid>
    <pubDate>Wed, 29 Jul 2026 00:00:00 GMT</pubDate>
    <description>Перевод статьи на zh-Hant</description>
    <category>liquidations</category>
    <category>cascade</category>
    <category>trading signal</category>
    <category>defi</category>
    <category>perpetuals</category>
    <category>order book</category>
    <category>market microstructure</category>
    <category>crypto</category>
    <enclosure url="https://marketmaker.cc/images/blog/liquidation-cascades-trading-signal.webp" type="image/webp" />
  </item>
  <item>
    <title>Uniswap v3 LP 策略與對沖</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/uniswap-v3-lp-strategies-hedging/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/uniswap-v3-lp-strategies-hedging/</guid>
    <pubDate>Tue, 28 Jul 2026 00:00:00 GMT</pubDate>
    <description>Перевод статьи на zh-Hant</description>
    <category>uniswap</category>
    <category>defi</category>
    <category>market making</category>
    <category>delta hedging</category>
    <category>lvr</category>
    <category>jit liquidity</category>
    <category>quant</category>
    <category>ethereum</category>
    <enclosure url="https://marketmaker.cc/images/blog/uniswap-v3-lp-strategies-hedging.webp" type="image/webp" />
  </item>
  <item>
    <title>MEV 供應鏈：PBS、MEV-Boost 與區塊市場</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/mev-supply-chain-pbs-mevboost/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/mev-supply-chain-pbs-mevboost/</guid>
    <pubDate>Mon, 27 Jul 2026 00:00:00 GMT</pubDate>
    <description>Перевод статьи на zh-Hant</description>
    <category>mev</category>
    <category>pbs</category>
    <category>mev-boost</category>
    <category>flashbots</category>
    <category>jito</category>
    <category>order flow auction</category>
    <category>defi</category>
    <category>ethereum</category>
    <enclosure url="https://marketmaker.cc/images/blog/mev-supply-chain-pbs-mevboost.webp" type="image/webp" />
  </item>
  <item>
    <title>鏈上套利：原子循環、閃電貸與你必須贏得的拍賣</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/onchain-arbitrage-atomic-flash-loans/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/onchain-arbitrage-atomic-flash-loans/</guid>
    <pubDate>Sun, 26 Jul 2026 00:00:00 GMT</pubDate>
    <description>找到鏈上套利循環是較容易的一半。困難的一半是贏得密封式競價優先拍賣，與所有找到相同循環的其他搜尋者競爭。對原子性、閃電貸、跟單交易的技術性剖析，以及套利利潤實際得以保留的地方——包括非原子的 CEX-DEX。</description>
    <category>arbitrage</category>
    <category>flash loans</category>
    <category>mev</category>
    <category>defi</category>
    <category>backrunning</category>
    <category>cex-dex</category>
    <category>ethereum</category>
    <category>atomicity</category>
    <enclosure url="https://marketmaker.cc/images/blog/onchain-arbitrage-atomic-flash-loans.webp" type="image/webp" />
  </item>
  <item>
    <title>無常損失與 LVR：LP 獲利能力的真實數學</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/impermanent-loss-lvr-lp-profitability/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/impermanent-loss-lvr-lp-profitability/</guid>
    <pubDate>Sat, 25 Jul 2026 00:00:00 GMT</pubDate>
    <description>無常損失、集中流動性槓桿以及從第一性原理推導出的再平衡損失。閉合形式、σ²/8 結果、每區塊標記價差，以及 LP 實際擊敗持有情況下的精確費用與波動性條件。</description>
    <category>defi</category>
    <category>amm</category>
    <category>impermanent loss</category>
    <category>lvr</category>
    <category>uniswap</category>
    <category>market making</category>
    <category>volatility</category>
    <category>quant</category>
    <enclosure url="https://marketmaker.cc/images/blog/impermanent-loss-lvr-lp-profitability.webp" type="image/webp" />
  </item>
  <item>
    <title>切片內部:排程器與交易所之間的子訂單戰術</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/child-order-execution-tactics/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/child-order-execution-tactics/</guid>
    <pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate>
    <description>Almgren-Chriss 與 VWAP 只負責設定切片預算——執行的勝負卻在戰術層決出。升級計時器、maker-taker 盈虧平衡數學、Binance/OKX/CME 上 amend 與 cancel-replace 的佇列語義、冰山訂單的反訊號洩露,以及一套 Python 逐切片狀態機。</description>
    <category>執行</category>
    <category>子訂單</category>
    <category>訂單戰術</category>
    <category>佇列位置</category>
    <category>maker-taker</category>
    <category>冰山訂單</category>
    <category>微觀結構</category>
    <category>python</category>
    <enclosure url="https://marketmaker.cc/images/blog/child-order-execution-tactics.webp" type="image/webp" />
  </item>
  <item>
    <title>滑點曲線,而非滑點常數:經得起實盤檢驗的成本模型</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/slippage-cost-models-backtest/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/slippage-cost-models-backtest/</guid>
    <pubDate>Thu, 23 Jul 2026 00:00:00 GMT</pubDate>
    <description>用依賴於規模、波動率和流動性的成本曲線取代恆定的 bps 滑點:平方根定律、從 TCA 成交或公開資料擬合曲線、市場狀態壓力乘數,以及為什麼你的策略排行榜會重新洗牌。</description>
    <category>滑點</category>
    <category>交易成本</category>
    <category>市場衝擊</category>
    <category>回測</category>
    <category>執行</category>
    <category>tca</category>
    <category>平方根定律</category>
    <category>成本模型</category>
    <enclosure url="https://marketmaker.cc/images/blog/slippage-cost-models-backtest.webp" type="image/webp" />
  </item>
  <item>
    <title>加密市場的智慧訂單路由：一筆訂單，十二個場所，沒有 NBBO</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/smart-order-routing-crypto/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/smart-order-routing-crypto/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
    <description>為什麼加密 SOR 比股票路由更難：沒有合併行情、幻影流動性、預注資本。一套帶數學與 Python 的凸路由最佳化、maker 感知的戰術，以及來自你自己 TCA 的分場所 markout 排行榜。</description>
    <category>執行</category>
    <category>智慧訂單路由</category>
    <category>市場微觀結構</category>
    <category>流動性碎片化</category>
    <category>crypto</category>
    <category>quant</category>
    <category>python</category>
    <category>tca</category>
    <enclosure url="https://marketmaker.cc/images/blog/smart-order-routing-crypto.webp" type="image/webp" />
  </item>
  <item>
    <title>做市與吃單的抉擇：費率檔位、返佣與跨越價差的真實成本</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/maker-taker-fees-rebates-execution/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/maker-taker-fees-rebates-execution/</guid>
    <pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate>
    <description>做市/吃單之爭本質是一筆披著費率外衣的逆向選擇交易。本文通過Glosten-Milgrom模型推導盈虧平衡數學、剖析真實加密貨幣費率檔位與代幣折扣、回顧返佣挖礦的歷史、分析佇列位置對返佣價值的影響，並講解如何在回測中建模分級費率。</description>
    <category>做市吃單</category>
    <category>交易費用</category>
    <category>返佣</category>
    <category>逆向選擇</category>
    <category>做市</category>
    <category>執行</category>
    <category>市場微觀結構</category>
    <category>回測</category>
    <enclosure url="https://marketmaker.cc/images/blog/maker-taker-fees-rebates-execution.webp" type="image/webp" />
  </item>
  <item>
    <title>執行落差與自建TCA：衡量執行到底讓你付出了多少代價</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/implementation-shortfall-tca-execution/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/implementation-shortfall-tca-execution/</guid>
    <pubDate>Mon, 20 Jul 2026 00:00:00 GMT</pubDate>
    <description>Perold的執行落差理論,被落地成一套針對加密貨幣交易機器人的可執行TCA流水線:基於到達價的分解、t+1秒/10秒/60秒的markout曲線、約200行Python程式碼即可處理你自己的成交記錄,並將結果回灌到你的回測成本模型中。</description>
    <category>執行</category>
    <category>TCA</category>
    <category>執行落差</category>
    <category>markouts</category>
    <category>交易成本</category>
    <category>演算法交易</category>
    <category>回測</category>
    <category>市場微觀結構</category>
    <enclosure url="https://marketmaker.cc/images/blog/implementation-shortfall-tca-execution.webp" type="image/webp" />
  </item>
  <item>
    <title>鏈上清算：Aave 與 Compound 的機制，以及圍繞它們的機器人生意</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/onchain-liquidations-aave-compound/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/onchain-liquidations-aave-compound/</guid>
    <pubDate>Sun, 19 Jul 2026 00:00:00 GMT</pubDate>
    <description>Aave v3 和 Compound 的清算究竟是如何運作的：健康因子的數學原理、清算折扣係數、清算獎勵、Chainlink 預言機觸發條件，以及在 2026 年營運一個閃電貸清算機器人需要什麼。</description>
    <category>DeFi</category>
    <category>清算</category>
    <category>Aave</category>
    <category>Compound</category>
    <category>MEV</category>
    <category>Chainlink</category>
    <category>閃電貸</category>
    <category>加密貨幣</category>
    <enclosure url="https://marketmaker.cc/images/blog/onchain-liquidations-aave-compound.webp" type="image/webp" />
  </item>
  <item>
    <title>面向量化的 Uniswap v3：從第一性原理理解集中流動性與 tick 數學</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/uniswap-v3-concentrated-liquidity-quants/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/uniswap-v3-concentrated-liquidity-quants/</guid>
    <pubDate>Sat, 18 Jul 2026 00:00:00 GMT</pubDate>
    <description>Uniswap v3 在底層究竟是如何運作的：虛擬儲備、流動性 L、sqrtPriceX96、tick 數學、feeGrowthInside，以及為什麼一個 LP 頭寸本質上是一份賣出波動率的損益結構。附帶精確公式與完整的數值算例。</description>
    <category>Uniswap</category>
    <category>DeFi</category>
    <category>AMM</category>
    <category>集中流動性</category>
    <category>做市</category>
    <category>LVR</category>
    <category>量化</category>
    <category>以太坊</category>
    <enclosure url="https://marketmaker.cc/images/blog/uniswap-v3-concentrated-liquidity-quants.webp" type="image/webp" />
  </item>
  <item>
    <title>MEV 剖析：三明治攻擊、搶先交易與記憶體池的黑暗森林</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/mev-sandwich-frontrunning-mempool/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/mev-sandwich-frontrunning-mempool/</guid>
    <pubDate>Fri, 17 Jul 2026 00:00:00 GMT</pubDate>
    <description>你傳送到公共 RPC 的每一筆 swap，都是一個沒人有義務遵守的可見限價單。本文技術性拆解 MEV 提取的運作機制：記憶體池排序、Uniswap v2 三明治攻擊數學、PGA、Flashbots，以及交易者的防禦手段。</description>
    <category>mev</category>
    <category>三明治攻擊</category>
    <category>搶先交易</category>
    <category>mempool</category>
    <category>flashbots</category>
    <category>DeFi</category>
    <category>ethereum</category>
    <enclosure url="https://marketmaker.cc/images/blog/mev-sandwich-frontrunning-mempool.webp" type="image/webp" />
  </item>
  <item>
    <title>成交模擬：從收盤價幻想到佇列感知現實的階梯</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/fill-simulation-partial-fills-backtest/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/fill-simulation-partial-fills-backtest/</guid>
    <pubDate>Thu, 16 Jul 2026 00:00:00 GMT</pubDate>
    <description>成交模擬保真度的五級階梯——從收盤價成交到機率化佇列位置模型。將部分成交建模為狀態機，把限價單成交機率的上下界作為 PnL 區間，並給出對照實盤成交的校準閉環。</description>
    <category>成交模擬</category>
    <category>回測</category>
    <category>限價單</category>
    <category>佇列位置</category>
    <category>部分成交</category>
    <category>做市</category>
    <category>執行</category>
    <category>市場微觀結構</category>
    <enclosure url="https://marketmaker.cc/images/blog/fill-simulation-partial-fills-backtest.webp" type="image/webp" />
  </item>
  <item>
    <title>TWAP、VWAP與POV對比：如何選擇執行基準（以及何時它們會對你撒謊）</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/twap-vwap-pov-execution-algorithms/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/twap-vwap-pov-execution-algorithms/</guid>
    <pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate>
    <description>TWAP、VWAP和POV本質上都是對成交量預測的一次押注。我們剖析每種調度演算法背後隱藏的假設，構建加密貨幣日內成交量曲線，並在回放的L2數據上對三者進行正面對比測試。</description>
    <category>execution</category>
    <category>twap</category>
    <category>vwap</category>
    <category>pov</category>
    <category>implementation-shortfall</category>
    <category>volume-curve</category>
    <category>algotrading</category>
    <category>backtest</category>
    <enclosure url="https://marketmaker.cc/images/blog/twap-vwap-pov-execution-algorithms.webp" type="image/webp" />
  </item>
  <item>
    <title>去掉玄乎其辭的 Almgren-Chriss：一個下午就能實現的最優執行模型</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/almgren-chriss-optimal-execution/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/almgren-chriss-optimal-execution/</guid>
    <pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate>
    <description>Almgren-Chriss 最優執行模型的完整推導：線性衝擊、sinh/cosh 交易軌跡、有效前沿，以及從 Binance L2 和成交資料校準 eta、gamma、sigma 的可執行 Python 程式碼。</description>
    <category>執行</category>
    <category>Almgren-Chriss</category>
    <category>市場衝擊</category>
    <category>最優執行</category>
    <category>TWAP</category>
    <category>quant</category>
    <category>python</category>
    <category>加密貨幣</category>
    <enclosure url="https://marketmaker.cc/images/blog/almgren-chriss-optimal-execution.webp" type="image/webp" />
  </item>
  <item>
    <title>波動率目標化與基於GARCH預測的交易</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/volatility-targeting-garch-strategy/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/volatility-targeting-garch-strategy/</guid>
    <pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate>
    <description>GARCH波動率預測只有在能改善交易決策時才有價值。我們構建了一個波動率目標化的加密貨幣策略，誠實地評估預測質量，並在walk-forward回測中將GARCH與已實現波動率和EWMA基準進行比較。</description>
    <category>volatility</category>
    <category>GARCH</category>
    <category>volatility-targeting</category>
    <category>backtesting</category>
    <category>risk</category>
    <category>crypto</category>
    <category>algorithmic-trading</category>
    <enclosure url="https://marketmaker.cc/images/blog/volatility-targeting-garch-strategy.webp" type="image/webp" />
  </item>
  <item>
    <title>DCC-GARCH：配對交易與組合風險的動態相關性</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/dcc-garch-dynamic-correlation-crypto/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/dcc-garch-dynamic-correlation-crypto/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 GMT</pubDate>
    <description>加密貨幣的相關性並非恆定——每次下跌都會飆升至接近1。DCC-GARCH對時變相關性矩陣建模，為配對交易提供動態對沖比率，並給出誠實的時變組合風險。</description>
    <category>volatility</category>
    <category>GARCH</category>
    <category>DCC</category>
    <category>correlation</category>
    <category>portfolio</category>
    <category>pairs-trading</category>
    <category>crypto</category>
    <enclosure url="https://marketmaker.cc/images/blog/dcc-garch-dynamic-correlation-crypto.webp" type="image/webp" />
  </item>
  <item>
    <title>非對稱與厚尾GARCH：EGARCH、GJR與Student-t</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/asymmetric-garch-crypto-leverage/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/asymmetric-garch-crypto-leverage/</guid>
    <pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate>
    <description>普通GARCH(1,1)對好訊息和壞訊息一視同仁，並假設衝擊服從高斯分佈。EGARCH、GJR-GARCH以及Student-t/偏態-t新息修正了這兩個缺陷，為加密資產提供更真實的VaR和預期損失估計。</description>
    <category>volatility</category>
    <category>GARCH</category>
    <category>EGARCH</category>
    <category>risk</category>
    <category>VaR</category>
    <category>crypto</category>
    <category>algorithmic-trading</category>
    <enclosure url="https://marketmaker.cc/images/blog/asymmetric-garch-crypto-leverage.webp" type="image/webp" />
  </item>
  <item>
    <title>GARCH(1,1)：預測加密貨幣波動率</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/garch-volatility-forecasting-crypto/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/garch-volatility-forecasting-crypto/</guid>
    <pubDate>Fri, 10 Jul 2026 00:00:00 GMT</pubDate>
    <description>GARCH(1,1)模型如何捕捉加密貨幣中的波動率聚集現象，如何使用arch庫通過最大似然法擬合該模型，以及如何將條件方差預測轉化為倉位規模和動態止損。</description>
    <category>volatility</category>
    <category>GARCH</category>
    <category>risk</category>
    <category>forecasting</category>
    <category>crypto</category>
    <category>algorithmic-trading</category>
    <enclosure url="https://marketmaker.cc/images/blog/garch-volatility-forecasting-crypto.webp" type="image/webp" />
  </item>
  <item>
    <title>誠實的負面結果：數萬次回測、五大主流幣，沒有穩健優勢</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/honest-negative-no-robust-edge/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/honest-negative-no-robust-edge/</guid>
    <pubDate>Thu, 09 Jul 2026 00:00:00 GMT</pubDate>
    <description>搜尋與過擬合這條主線的收官之作，以一個負面結果收尾——而這正是正確的結果。ETHUSDT 上的單品種雙時間框架搜尋找到了一個樣本外 +16.35%、未觸碰留出視窗 +2.62% 的配置；Deflated Sharpe Ratio 在計入約 37,000 次試驗後把它折損到 0.00。對五個主流幣（ETH/BTC/SOL/BNB/XRP，每個約 1.18M 根 1 分鐘 K 線）按樣本外中位數選擇的跨品種檢驗則徹底終結了它：雙時間框架 DSR 0.24 / PBO 0.264，三時間框架 DSR 0.14 / PBO 0.327——兩者都沒過閘門。冠軍只在 5 個品種中的 1 個上盈利，其餘全部為負。這正是反過擬合裝置存在的意義：阻止你把噪聲中的最優當成 alpha 拿去上線。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>過擬合</category>
    <category>折損夏普比率</category>
    <category>PBO</category>
    <category>負面結果</category>
    <category>驗證</category>
    <enclosure url="https://marketmaker.cc/images/blog/honest-negative-no-robust-edge.webp" type="image/webp" />
  </item>
  <item>
    <title>證明多時間框架回測中沒有look-ahead：擾動未來，證明過去看不到它</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/proving-no-lookahead-multi-timeframe/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/proving-no-lookahead-multi-timeframe/</guid>
    <pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate>
    <description>多時間框架回測會通過一根仍在形成中、尚不存在最終收盤價的高時間框架K線洩露未來資訊。僅靠程式碼審查無法建立信心——必須實際測試。我們精確復現了實盤機器人的closed-bar規則，然後用一個“未來偏移探針”證明沒有資訊洩露：擾動每一根未來K線，並斷言每一個過去的訊號和交易都逐位不變。25/25項一致性檢查全部通過，而且這個探針確實有效（並非一個永遠不會失敗的空測試）。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>look-ahead偏差</category>
    <category>多時間框架</category>
    <category>資料洩露</category>
    <category>驗證</category>
    <enclosure url="https://marketmaker.cc/images/blog/proving-no-lookahead-multi-timeframe.webp" type="image/webp" />
  </item>
  <item>
    <title>GPU 何時才划算：參數掃描的屋頂線，一個 167x 頭條其實是 27x 演算法 × 6.2x 硬體</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/when-gpu-pays-off-sweep-roofline/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/when-gpu-pays-off-sweep-roofline/</guid>
    <pubDate>Tue, 07 Jul 2026 00:00:00 GMT</pubDate>
    <description>GPU 相對 CPU 的領先幅度隨批次大小增長——在我們的多時間框架指標預計算上，從每次呼叫一個組合時的 54.5x 一路升到 61 個組合時的 359.6x——因為一次小規模掃描無法攤薄核心啟動和傳輸開銷。我們把一個 167x 的頭條數字拆解為一個同樣惠及 CPU 的 27x 演算法收益，乘以一個 6.2x 的硬體收益，指出 GPU 相對最佳 CPU 的真實領先在單時間框架下只有 3.2x、在多時間框架下只有 6.2x，並給出一份決策指南：一次掃描要寬到什麼程度，才值得為 GPU 掏錢。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>效能</category>
    <category>gpu</category>
    <category>屋頂線</category>
    <category>最佳化</category>
    <enclosure url="https://marketmaker.cc/images/blog/when-gpu-pays-off-sweep-roofline.webp" type="image/webp" />
  </item>
  <item>
    <title>GPU 精度陷阱：Apple Metal 上的 fp32 回測如何悄無聲息地返回垃圾結果</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/gpu-precision-trap-fp32-backtest/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/gpu-precision-trap-fp32-backtest/</guid>
    <pubDate>Mon, 06 Jul 2026 00:00:00 GMT</pubDate>
    <description>Apple 的 Metal GPU 沒有 float64。把一個向量化回測天真地移植過去，那個誘人的字首和 WMA 就會在 fp32 下溢位——最大相對誤差 211 倍——但它依然能跑完全程，還返回看起來說得過去的數字。修復的辦法不是提高精度，而是換一種公式：直接的視窗卷積，fp32 下精度可達 8×10⁻⁷，比單執行緒 numba 快 55.9 倍。這篇文章講清楚陷阱本身、背後的算術，以及如何證明自己沒有掉進去。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>gpu</category>
    <category>蘋果晶片</category>
    <category>浮點數</category>
    <category>數值穩定性</category>
    <category>mlx</category>
    <enclosure url="https://marketmaker.cc/images/blog/gpu-precision-trap-fp32-backtest.webp" type="image/webp" />
  </item>
  <item>
    <title>保真度關卡：由粗到細的回測會更快地愚弄你——除非廉價代理的排序方式與昂貴評估一致</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/drill-down-multifidelity-fidelity-gate/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/drill-down-multifidelity-fidelity-gate/</guid>
    <pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate>
    <description>下鑽式/多保真度搜索（ASHA、逐次減半、Hyperband）以低成本方式篩選成千上萬個配置，只將存活者提升到昂貴的完整評估階段。這是一種真實的加速——但如果低保真度的排序與高保真度的排序不一致，它就會悄無聲息地崩潰。我們測量了折數-排序相關性：在只用一折時，Spearman ρ 可能低至 0.03（排序幾乎是隨機的），隨著折數累積，逐步升高到 0.43、0.67、0.78、0.91。修復方法是設定一道強制性關卡——先測量 ρ(廉價, 完整)，並自動將最低保真度提升到 ρ ≥ 0.5 的第一個檔位。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>多保真度</category>
    <category>Hyperband</category>
    <category>ASHA</category>
    <category>逐次減半</category>
    <category>過擬合</category>
    <category>驗證</category>
    <enclosure url="https://marketmaker.cc/images/blog/drill-down-multifidelity-fidelity-gate.webp" type="image/webp" />
  </item>
  <item>
    <title>隨機搜尋 vs 智慧搜尋：交叉點在評估成本，而非演算法本身</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/search-method-crossover-eval-cost/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/search-method-crossover-eval-cost/</guid>
    <pubDate>Sat, 04 Jul 2026 00:00:00 GMT</pubDate>
    <description>當單次回測很便宜時，簡單的打亂 Sobol 序列在原始吞吐量上碾壓所有&quot;智慧&quot;取樣器——TPE、CMA-ES、ASHA 要繳納 Python ask/tell 稅，導致速度下降 20 倍，在相同牆鍾時間內評估的點數遠遠更少，因而落敗。讓每次評估變得昂貴（多時間框架 + walk-forward 折）之後，交叉點就翻轉了。我們測量了這兩種情形，以及為什麼摺疊排名保真度（ρ@1 從 0.03 升至 0.43）是剪枝能否奏效的前提條件。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>超參數最佳化</category>
    <category>貝葉斯最佳化</category>
    <category>walk-forward</category>
    <category>過擬合</category>
    <enclosure url="https://marketmaker.cc/images/blog/search-method-crossover-eval-cost.webp" type="image/webp" />
  </item>
  <item>
    <title>雙軸參數空間：為什麼你的大部分參數掃描幾乎是免費的</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/two-axis-parameter-space/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/two-axis-parameter-space/</guid>
    <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
    <description>並非所有參數的搜尋成本都相同。一個策略的參數會分裂為一條昂貴軸（指標——需要在整個序列上重新計算）和一條廉價軸（決策閾值——對預計算訊號做一次 O(n) 遍歷）。由於指標對閾值不變，你只需計算一次，就能以約 5,600 cfg/s 的速度掃描數千種閾值配置——比每種配置都重新計算大約便宜 1,600 倍。這是對維度災難的一次重新定價。</description>
    <category>量化交易</category>
    <category>回測</category>
    <category>參數搜尋</category>
    <category>最佳化</category>
    <category>快取</category>
    <category>維度災難</category>
    <enclosure url="https://marketmaker.cc/images/blog/two-axis-parameter-space.webp" type="image/webp" />
  </item>
  <item>
    <title>框架稅：當你的回測庫比手寫的 pandas 迴圈還慢</title>
    <link>https://marketmaker.cc/zh-Hant/blog/post/framework-tax-event-driven-vs-vectorized/</link>
    <guid isPermaLink="true">https://marketmaker.cc/zh-Hant/blog/post/framework-tax-event-driven-vs-vectorized/</guid>
    <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
    <description>我們在同一個參數掃描任務上對八個回測引擎進行了基準測試——15萬根K線、80個HMA交叉組合、交易筆數嚴格鎖定在2707筆。兩個最受歡迎的事件驅動框架的速度居然比手寫的 pandas 迴圈還慢，而一個向量化/編譯型引擎完成同樣的工作快了約13000倍。這是一篇關於主流庫從未被設計用來攤銷的逐K線開銷的研究。</description>
    <category>演算法交易</category>
    <category>回測</category>
    <category>效能</category>
    <category>參數搜尋</category>
    <category>向量化</category>
    <category>基準測試</category>
    <enclosure url="https://marketmaker.cc/images/blog/framework-tax-event-driven-vs-vectorized.webp" type="image/webp" />
  </item>
</channel>
</rss>