Alex Nedyalkov·Jul 11Verification Is the Bottleneck: Why “Master Analyst” Needs a Probabilistic VerifierA technical breakdown for anyone building agentic systems for market analysis
Alex Nedyalkov·Jul 5Moving Beyond Brute Force in Time Series ForecastingIn my recent work, I spend a lot of time thinking about how to get the most out of forecasting models for European electricity markets…
Alex Nedyalkov·May 16The Bitter Lesson of 2026: The Great Convergence of Time Series AIFrom the ICLR 2025 scaling laws to the 2.5B parameter Toto-2, the verdict is in: scale is winning, but the specialists aren’t going down…
Alex Nedyalkov·Apr 28Why Preprocessing Details Fade to Noise: Scaling Lessons from Amazon’s Chronos-2Last month I argued that TOTO and Chronos-2 represent a philosophical split in time-series foundation models — the specialist versus the…
Alex Nedyalkov·Apr 5Modern Forecasting Has Outgrown Classical BenchmarksA recent position paper, “Seeking SOTA: Time-Series Forecasting Must Adopt Taxonomy-Specific Evaluation to Dispel Illusory Gains” (Saqur et…
Alex Nedyalkov·Mar 19TOTO vs Chronos-2: Why TOTO’s Mixture Head Might Outperform Chronos-2 for Volatile Price DataChronos-2 vs. TOTO: The Battle for Multivariate Supremacy
Alex Nedyalkov·Mar 5From “XGB Everything” to Agentic Scaffolds: The Great Forecasting Pivot of 2025–2026The question for 2026 is no longer, “Which model is most accurate?” but rather, “How robust is your scaffold?”
Alex Nedyalkov·Feb 19When “More Data” Isn’t Better: Two Shifts That Are Quietly Reshaping Machine LearningFor years, two assumptions have quietly guided most machine learning practice:A response icon1A response icon1
Alex Nedyalkov·Nov 24, 2025Beyond Gemini 3: Google’s “Synapse” and the Quiet Revolution in Time Series ForecastingWhile the industry scrambles to benchmark the newly released Gemini 3, a more subtle but equally critical shift occurred two weeks ago…
Alex Nedyalkov·Nov 11, 2025Your Time Series AI is Biased (And That’s a Good Thing)Chronos 2’s Hidden Trade-Off: How Architectural Bias Cut My Forecasting Error by 58%