https://nosible.com/blog/an-embedding-based-approach-to-trade-and-economic-policy-uncertainty 2026-06-17 https://nosible.com/images/2026/06/nosible-tpu-daily.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty Rebuild the Fed’s Trade Policy Uncertainty index from 14.9 million NOSIBLE World events using embeddings, then extend the method to broader policy risk. https://nosible.com/images/2026/06/nosible-tpu-threeway.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty Three monthly measures of trade policy uncertainty, rebased to a common mean and aligned on the same dates, the published TPU, the Baker-Bloom-Davis trade-policy EPU, and NOSIBLE-TPU. All three track each other and peak together in April 2025. Against the published TPU the two official measures agree at 0.96 on levels and 0.70 on changes, while NOSIBLE matches it at 0.87 and 0.82, so on month-to-month changes it tracks the benchmark more closely than the two official measures track each other. https://nosible.com/images/2026/06/nosible-tpu-polarity.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty The daily net polarity of trade-policy coverage, scored from -1 when coverage is resolution-framed to +1 when it is uncertainty-framed, with a state-dependent smoothing overlay. The line runs positive during escalations, +0.59 at the 2018 China tariffs and +0.48 on Liberation Day, sits near zero through the quiet stretch of 2021, and turns negative when a dispute resolves, such as the January 2020 Phase One deal. This signed direction is what the published level-only index cannot show, telling a resolved story apart from a simply quiet news week. https://nosible.com/images/2026/06/nosible-tpu-tariffs.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty The realized US tariff rate, quarterly, plotted against all three trade-policy-uncertainty indices, 2014 to 2026. The tariff rate edges up through the 2018-19 trade war, from 1.4% to 3%, then explodes from 2.6% to 12.8% in 2025, the largest move since the 1970s. The news indices lead the realized rate. Each index's level lines up with the next quarter's change in the tariff rate at 0.61 for TPU, 0.59 for BBD and 0.66 for NOSIBLE. Higher uncertainty tends to be followed by higher tariffs. https://nosible.com/images/2026/06/nosible-epu-us.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty NOSIBLE-EPU against the published US Economic Policy Uncertainty index, monthly and rebased, with a nine-lever version that drops the shutdown lever shown as a diagnostic line. NOSIBLE tracks the published index at 0.77 on levels and 0.53 on changes. Both rank April 2025 the highest month of the period, and both place the COVID shock of spring 2020 near the top. The two NOSIBLE lines agree everywhere except 2020, which shows that the entire COVID gap comes down to the single shutdown lever. https://nosible.com/images/2026/06/nosible-epu-categories.png An Embedding-Based Approach to Trade and Economic Policy Uncertainty The national-security and healthcare categories of Economic Policy Uncertainty, NOSIBLE against the published categorical series, monthly. National security tracks the published series at 0.83 on levels and 0.59 on changes, with both ranking April 2025 first. Healthcare is looser at 0.73 and 0.44 but agrees on the periods that dominate the series, the COVID spring of 2020 and the 2025 health-policy upheaval. Both categories come from the same sentence set as the headline index, with no extra word lists. https://nosible.com/blog/can-faceted-search-at-web-scale-self-organize 2025-10-16 https://nosible.com/blog/illustrations/vortex.png Can Faceted Search at Web-Scale Self Organize? See how adaptive named-entity tagging lets a web-scale search index organize itself into useful facets without relying on a fixed taxonomy. https://nosible.com/blog/ensemble-and-distil 2024-02-06 https://nosible.com/blog/illustrations/cube.png A Pattern for Scaling the Value Proposition of LLMs: Ensemble and Distil 🚀 Train a simple regression on sentence embeddings to distil an LLM ensemble—and outperform GPT-4 on financial sentiment classification. https://nosible.com/blog/fast-enough-to-matter-productionizing-tiny-transformers-for-signal-extraction 2025-12-12 https://nosible.com/blog/illustrations/the-sprinter.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 Fine-tune Qwen3 0.6B with active learning to match GPT-5.1 on financial sentiment, with open models, datasets and training code. https://nosible.com/images/2025/12/financial_sentiment_accuracy_on_phrasebank_dataset.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 Financial sentiment accuracy comparison on PhraseBank dataset https://nosible.com/images/2025/12/financial_sentiment_accuracy_on_nosible_fin_sent_dataset.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 Financial sentiment accuracy comparison on NOSIBLE financial sentiment dataset https://nosible.com/images/2025/12/financial_sentiment_accuracy_vs_cost_on_phrasebank_dataset.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 Accuracy vs Cost plot: Financial sentiment accuracy on Financial PhraseBank vs Cost https://nosible.com/images/2025/12/financial_sentiment_accuracy_vs_cost_on_nosible_dataset.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 Accuracy vs Cost plot: Financial sentiment accuracy on NOSIBLE financial sentiment dataset vs Cost https://nosible.com/images/2025/12/openrouter-qwen3-embedding-8b-leaderboard.png Matching GPT-5.1 at Financial Sentiment with Active Learning and Qwen3 OpenRouter leaderboard showing NOSIBLE ranking 6th for Qwen3-Embedding-8B usage after the project https://nosible.com/blog/introducing-cybernaut-1-agentic-search-with-mcts 2025-08-26 https://nosible.com/blog/illustrations/cyber.png Introducing Cybernaut-1: Agentic Search using MCTS Meet Cybernaut-1, NOSIBLE’s agentic search system combining hybrid retrieval with LLM-guided Monte Carlo Tree Search. https://nosible.com/blog/news-sentiment-showdown-who-checks-vibes-best 2024-01-28 https://nosible.com/blog/illustrations/eye.png News Sentiment Showdown: Who Checks Vibes Best? Compare TextBlob, VADER, FinBERT, Gemini, GPT-3.5 and GPT-4 on 10,368 labelled financial news stories, including the dataset and code. https://nosible.com/blog/point-in-time-knowledge-graphs-over-named-entities 2026-07-16T10:31:25.000Z https://nosible.com/images/2026/07/kg-hero-decade.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World Build point-in-time knowledge graphs from dated NOSIBLE World events using named entities and lift-scored co-mentions—without lookahead bias. https://nosible.com/images/2026/07/kg-network-nvda.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World NVIDIA's 2024 co-mention network. Nodes are entities co-mentioned with NVIDIA; lines join entities that co-occur with each other, so related nodes cluster. Brand green marks products, lighter green peers, mist people, muted green organizations, an outline places; size is association strength. The supply chain, the accelerator platform and the AI cloud each form a group. Built from news on or before 31 December 2024. https://nosible.com/images/2026/07/kg-evolution-matrix.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World The NVIDIA product line-up over time, showing 40 of the 332 products NVIDIA is co-mentioned with. Dot size is that year's association strength; colour is how the product's share of NVIDIA's coverage changed from the year before, green for growing, red for shrinking, neutral for flat. RTX enters in 2018, the A100 in 2020, and Blackwell, HBM and the H200 arrive in 2024. https://nosible.com/images/2026/07/kg-network-aapl.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World Apple's 2024 co-mention network, drawn the same way. Its product line sits at the centre, ringed by peers such as Samsung and Foxconn, the analysts and executives who cover it (Tim Cook, Mark Gurman, Ming-Chi Kuo), and the platforms it competes and partners with (Google, Microsoft, Spotify). Built from news on or before 31 December 2024. https://nosible.com/images/2026/07/kg-aapl-products-matrix.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World Apple's product line-up over time, 40 of 274 products, drawn the same way. The iPhone, iPad and App Store run throughout; the Apple Watch and AirPods arrive in the mid 2010s, the iPhone X in 2018, AirPods Pro and the Pro Max line from 2019, then AirTag and Dynamic Island, and Vision Pro in 2025. https://nosible.com/images/2026/07/kg-network-ko.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World Coca-Cola's 2021 co-mention network, drawn the same way. The bottling system clusters on one side and the beverage competitors on another. The lower group is the backlash to Georgia's 2021 voting law, when Coca-Cola and Delta criticized it and Major League Baseball pulled its All-Star Game from Atlanta, so Georgia, Atlanta, Delta and the MLB appear together. Built from news on or before 31 December 2021. https://nosible.com/images/2026/07/kg-ko-people-matrix.png Point-in-Time Knowledge Graphs over Named Entities with NOSIBLE World Coca-Cola's associated people over time. Muhtar Kent is present from 2015 to 2018, James Quincey from 2016 on, and Cristiano Ronaldo shows up as a single spike in 2021, the year of his Euro 2020 press-conference moment. https://nosible.com/blog/rebuilding-the-geopolitical-risk-index-from-nosible-world 2026-06-06 https://nosible.com/images/2026/06/nosible-gpr-vs-published.png We Rebuilt the Geopolitical Risk Index with Nosible World Rebuild the geopolitical risk index from 13.2 million NOSIBLE World events and reproduce its global, country, country-pair and oil-risk signals. https://nosible.com/images/2026/06/nosible-gpr-by-country.png We Rebuilt the Geopolitical Risk Index with Nosible World Per-country signal, same-day and 12-month detrended, against the published country indices for Russia, Israel, Ukraine, Iran, the USA, and India https://nosible.com/images/2026/06/nosible-gpr-bilateral.png We Rebuilt the Geopolitical Risk Index with Nosible World Bilateral signal, same-day and 12-month detrended, for major country pairs against the published bilateral series https://nosible.com/images/2026/06/nosible-tariff-recovery.png We Rebuilt the Geopolitical Risk Index with Nosible World China and the China-USA pair, before and after folding in trade coercion. AI-GPR in amber, the Nosible baseline in grey, the Nosible version with tariffs in green https://nosible.com/images/2026/06/nosible-oil-gpr-vs-wti.png We Rebuilt the Geopolitical Risk Index with Nosible World Nosible Oil-GPR, 12-month detrend in green and same-day in blue, with the published academic Oil-GPR in amber and the WTI oil price in grey. After Iacoviello and Tong, 2026, Figure 4 https://nosible.com/images/2026/06/nosible-oil-gpr-by-region.png We Rebuilt the Geopolitical Risk Index with Nosible World Per-region oil-supply risk: the published academic version in amber, the Nosible version in green, by producer region, monthly and z-scored https://nosible.com/blog/the-contrastive-geometry-of-risk 2026-06-18 https://nosible.com/images/2026/06/walkthrough_step9_number_line.png Two Tricks for Turning Sentence Embeddings into Clean Features Turn OpenAI sentence embeddings into clean geographic and systemic-risk features with two deterministic, training-free scoring techniques. https://nosible.com/images/2026/06/walkthrough_step3_geo_raw.png Two Tricks for Turning Sentence Embeddings into Clean Features Strip plot of the input sentence's 60 raw cosines, 20 per bucket, colored local, national, global. The clouds are noisy and overlap in the middle. https://nosible.com/images/2026/06/walkthrough_step4_geo_gated.png Two Tricks for Turning Sentence Embeddings into Clean Features The same 60 cosines after gating. The background is squashed toward 0 and the strong global matches push toward 1. https://nosible.com/images/2026/06/walkthrough_step5_geo_scores.png Two Tricks for Turning Sentence Embeddings into Clean Features Bar chart of the three bucket scores: local 0.315, national 0.417, global 0.739. The argmax is global. https://nosible.com/images/2026/06/walkthrough_step7_scope_raw_2d.png Two Tricks for Turning Sentence Embeddings into Clean Features A 2D projection of the raw mirror-set embeddings. The idiosyncratic and systemic clouds overlap heavily. https://nosible.com/images/2026/06/walkthrough_step8_scope_neut_2d.png Two Tricks for Turning Sentence Embeddings into Clean Features The same sets after neutralization, projected onto the contrast direction. Idiosyncratic on the right, systemic on the left, cleanly separated. https://nosible.com/images/2026/06/walkthrough_step10_fusion.png Two Tricks for Turning Sentence Embeddings into Clean Features A 2D scatter of the verification events plus the running input, each labelled by name and positioned by geographic scale and scope. The running input sits in the global, systemic corner. https://nosible.com/blog/the-road-to-cybernaut-1 2025-08-20 https://nosible.com/blog/illustrations/track.png The Road to Cybernaut-1: Rebuilding Search for AI Why AI needs a purpose-built search engine, and how NOSIBLE rebuilt hybrid retrieval on the road to its Cybernaut-1 agentic search system. https://nosible.com/blog/turning-news-into-a-risk-on-risk-off-equity-signal 2026-06-16 https://nosible.com/images/2026/06/news-stress-overlay-hero.png Turning News into a Risk-On/Risk-Off Equity Signal Build a point-in-time market-stress signal from NOSIBLE World that cuts S&P 500 drawdown and transfers unchanged to the Nasdaq and Russell 2000. https://nosible.com/images/2026/06/news-stress-relevance-histogram.png Turning News into a Risk-On/Risk-Off Equity Signal Histogram of stress-relevance scores for real news events: the x-axis is each event's maximum cosine similarity to the 17 stress anchors, the green bars pile up near zero, and an amber dashed line marks the 0.30 relevance floor above which an event counts toward the daily reading https://nosible.com/images/2026/06/news-stress-calibration.png Turning News into a Risk-On/Risk-Off Equity Signal The calibration pipeline in four stacked panels, top to bottom: the raw daily attention share, a 7-day rolling mean, the trailing 252-day robust z-score, and the EWMA-smoothed z that is actually traded, with the de-risk threshold drawn on the lower panels https://nosible.com/images/2026/06/news-stress-episode-tariff-shock-2025.png Turning News into a Risk-On/Risk-Off Equity Signal The 2025 tariff shock in two stacked panels: on top the S&P 500 growth of $1 with the out-of-equities window shaded red, and below the lagged news-stress z crossing the exit line (z greater than 1.75, dashed) in February and falling back under the re-enter line (z less than 0.25, dotted) in late May https://nosible.com/images/2026/06/news-stress-signal-vs-vix.png Turning News into a Risk-On/Risk-Off Equity Signal The calibrated news-stress z (green, left axis) plotted against the VIX (amber, right axis) from 2010 to 2026, with the periods where the z sits above the de-risk threshold shaded red https://nosible.com/images/2026/06/news-stress-overlay-sp500.png Turning News into a Risk-On/Risk-Off Equity Signal S&P 500 news-stress overlay in three stacked panels, 2015 to 2026: log growth of $1 for buy-and-hold (white) versus the overlay (green), the drawdown path of each, and the lagged news-stress z with its exit and re-enter thresholds; red shading marks the days the overlay spent out of equities https://nosible.com/images/2026/06/news-stress-overlay-nasdaq.png Turning News into a Risk-On/Risk-Off Equity Signal Nasdaq Composite news-stress overlay versus buy-and-hold, 2015 to 2026: log growth of $1 (overlay in green, buy-and-hold in white), the drawdown of each, and the lagged news-stress signal, with the days out of equities shaded red https://nosible.com/images/2026/06/news-stress-overlay-russell.png Turning News into a Risk-On/Risk-Off Equity Signal Russell 2000 (IWM) news-stress overlay versus buy-and-hold, 2015 to 2026: log growth of $1 (overlay in green, buy-and-hold in white), the drawdown of each, and the lagged news-stress signal, with the days out of equities shaded red https://nosible.com/blog/using-vector-search-to-see-signals-in-company-news 2024-01-21 https://nosible.com/blog/illustrations/inspect.png Using Vector Search to See Signals in Company News Use vector search over company news to extract investment signals from a multi-terabyte, point-in-time corpus of more than 55 million embeddings.