How DevTrends Works

DevTrends tracks technologies across four signal types — developer community activity, job market demand, ecosystem health, and open-source momentum. These signals are collected from 9+ sources daily and combined into a single composite score from 0–100.

The goal is to answer three questions a developer actually cares about: Is this technology growing or shrinking? Is it in demand from employers? Is the community healthy?

We don't measure quality, suitability for a specific project, or personal preference. A high score means a technology is widely adopted, actively discussed, and currently hired for — not that it's the right tool for your situation.

We collect data from 9+ sources, organized into four signal categories:

Each technology gets a composite score from 0 to 100 using weights that adapt based on the technology's category and maturity. A programming language and a CSS framework shouldn't be weighted the same way — the language matters more in the job market, the framework matters more in community adoption.

Weight profiles by category

Maturity adjustments

New technologies (under 6 months of data) get boosted GitHub and community weights — because job market adoption lags behind developer interest by months. Mature technologies get boosted job and ecosystem weights, since those signals are more stable predictors of long-term relevance.

Score ranges

Every score includes a confidence grade (A–F) based on data completeness, sample size, recency, and source diversity. A score of 85 with grade A is more reliable than the same score with grade C — it means more sources agreed and the data is fresh.

Raw numbers aren't comparable across technologies — a framework with 50,000 GitHub stars and one with 500 stars aren't in the same league. We normalize each signal using z-score normalization, which measures how far a technology sits from the average across the full tracked population.

We then apply Bayesian smoothing to prevent low-sample technologies from gaming the rankings. Without it, a project with 10 stars gaining 10 more would score higher than one with 10,000 stars gaining 500 (100% vs 5% growth). Smoothing applies a confidence penalty proportional to sample size — the less data we have, the more we pull the score toward the population mean.

The composite score tells you where a technology stands today. Momentum tells you where it's heading. We use three exponential moving averages running in parallel to distinguish short-term noise from genuine directional change.

When short-term momentum contradicts long-term momentum — a 7-day spike against a 90-day decline, for example — the system flags a potential inflection point. These are worth investigating further before drawing conclusions.

Trend labels

Mention count alone is misleading. "React 19 is incredible" and "Why I'm abandoning React" are both mentions — but they carry opposite signals. Our sentiment engine combines lexicon-based analysis with tech-specific context rules to weight mentions by their actual signal direction.

Sentiment adjusts the community sub-score by up to ±15 points. High mention volume with predominantly negative sentiment will score meaningfully lower than moderate volume with positive sentiment.

Scores and charts show you what's happening. The AI insights explain why — drawing on the underlying data to surface patterns, flag anomalies, and provide context a raw number can't.

Each technology receives a human-readable status label derived from its score, momentum direction, and confidence grade combined:

Data collection and score recomputation run on a fixed schedule:

We'd rather tell you where this breaks than have you discover it the wrong way.

• •Popularity ≠ quality. A high score means a technology is widely used, discussed, and hired for. It says nothing about whether it's the right tool for your specific situation.
• •Enterprise blind spots. Technologies dominant in enterprises (Java, C#, SAP, Oracle) may score lower on community signals because enterprise discussions happen behind firewalls, not on Hacker News.
• •English-language bias. All data sources are English-language. Technologies with large non-English communities — particularly in China, Japan, and Eastern Europe — may be underrepresented.
• •AI-generated insights can be wrong. Insights are quality-checked but not fact-checked by humans. LLMs can misinterpret data or miss context. Use them as a starting point, not a final answer.
• •Job data lags reality. Job postings are a lagging indicator. By the time a technology shows up heavily in job listings, the early-adopter window has often already closed.
• •Sentiment accuracy is imperfect. Our tech-aware sentiment analysis is more accurate than generic approaches, but sarcasm detection and context understanding aren't solved problems. Ironic or nuanced community discussion may be misclassified.