About Almanac

We're cartographers, not prophets. We map the territory of possible futures with no stake in which one arrives.

What This Is

Almanac is an AI-powered forecasting system that tracks how software engineering is changing over the next 3-30 years. Every day, our pipeline scrapes 17+ data sources, filters hundreds of signals through tiered AI models, and produces a confidence-scored forecast. We track 55 specific, falsifiable predictions and update them daily using Bayesian likelihood ratio analysis with a 10-persona forecaster panel.

What This Is Not

We are not an oracle. We are not financial advisors. Our predictions are probabilistic estimates based on available evidence, and they are frequently wrong. The value isn't in any single prediction being correct — it's in the discipline of tracking, updating, and honestly reporting our accuracy over time.

How We're Different

Almanac
  • 55 SWE-specific predictions
  • Daily automated Bayesian updates
  • 10-persona panel with evidence trails
  • Free, open methodology
Metaculus
  • Broad topics (not SWE-focused)
  • Crowd-sourced human forecasters
  • Large resolved question library
  • Free, community-driven
Polymarket
  • Financial prediction markets
  • Real money at stake
  • Mostly politics/crypto/events
  • Not SWE career-focused

Methodology

01

Data Collection

Every day at 06:00 UTC, automated scrapers collect data from 17+ sources: Hacker News, arXiv, 5 RSS feeds (TechCrunch, Ars Technica, MIT Tech Review, The Verge, Wired), GitHub Trending, Stack Overflow, Semantic Scholar, FRED economic data, SEC EDGAR filings, Reddit, Lobsters, investor relations feeds, and X/Twitter. This yields 500+ raw items per day.

02

Signal Filtering

The AI model evaluates each item for relevance to the future of software engineering (3-30 year horizon). Items scoring below 0.5 recency-adjusted relevance are discarded. Typically 20-30 signals survive filtering. Each signal is classified by type: AI coding, job market, skills, regulation, tooling, or paradigm.

03

Narrative Synthesis

The top signals are synthesized into a coherent daily narrative. The AI is provided with yesterday's executive summary for continuity. It identifies the day's most significant developments and explains their 3-30 year implications.

04

Prediction Update

Each of our 55 standing predictions is evaluated against the day's evidence via Bayesian likelihood ratio analysis. A 10-persona forecaster panel votes independently. Confidence can move up to +/-5 percentage points per day. All movements are logged with evidence trails.

05

Publication

The report is committed to our private repository (with prediction deltas in the commit message), then pushed to the public site. The entire pipeline runs in under 10 minutes. Every report is timestamped and immutable.

The 10 Forecaster Personas

Each prediction is independently evaluated by 10 AI personas with different worldviews. The median likelihood ratio is used — no single perspective dominates.

🚀

Techno-Optimist

Believes AI will accelerate everything. Weights positive adoption signals heavily.

📊

Labor Economist

Focuses on employment data, wage trends, historical automation parallels.

🛡

Security Hawk

Highlights risks, vulnerabilities, and regulatory responses to AI-generated code.

Contrarian

Systematically challenges consensus. Asks 'what if the opposite happens?'

📈

Base Rate Empiricist

Anchors on historical base rates. Skeptical of 'this time is different' narratives.

💡

Startup Founder

Bullish on disruption speed. Weights funding/valuation signals. Believes incumbents are slow.

🏢

Enterprise Architect

Conservative on adoption timelines. Knows procurement cycles take 18-24 months. Demo ≠ deployed.

🌐

Open Source Advocate

Bullish on community-driven development. Believes open-source always wins long-term.

⚖️

Regulatory Watcher

Believes regulation is coming faster than industry expects. EU AI Act, liability trends.

🎓

Developer Educator

Focused on skills, bootcamps, CS enrollment. Believes talent pipeline adapts faster than pessimists predict.

Design Decisions

The numbers behind our methodology aren't arbitrary — here's why we chose them.

0.3 dampening

At 1.0, a single day's evidence from a noisy source causes multi-percentage-point swings that reverse the next day. At 0.3, it takes consistent evidence across several days to meaningfully shift a prediction — which matches how real-world trends work.

5pp max daily move

Prevents overreaction to single events. Even genuinely significant developments (a major acquisition, a breakthrough paper) need time for second-order effects to become clear. Big shifts should accumulate over weeks, not happen overnight.

10 personas, median LR

Using the median (not mean) of 10 likelihood ratios is robust to outlier personas. If the Techno-Optimist gives LR=5.0 and everyone else gives ~1.0, the median ignores the outlier. This prevents any single worldview from dominating.

[3%, 97%] bounds

Nothing is ever truly 0% or 100% — there's always a chance we're wrong about the question itself, or the world changes in ways nobody predicted. Hard bounds acknowledge irreducible uncertainty.

Try the Formula Yourself

This calculator runs the exact same Bayesian update formula our pipeline uses every day. Use the presets to see how source credibility and evidence strength interact.

Bayesian Update Calculator

Mirrors the exact formula used by our daily pipeline. Adjust sliders to see how evidence moves a prediction.

Presets:
50%
3%97%
1.5×
0.2× (strong against) 5.0× (strong for)
0.30
0.0 (no update) 1.0 (full Bayes)
1.0
0.5 (social/noise) 1.5 (research paper)
New probability
50%
+0.0 pp
No cap applied
Step-by-step math
Log-odds (prior): 0.000
Effective LR (LR^cred): 1.500
Log(eff. LR): 0.405
× Dampening: 0.122
New log-odds: 0.122
Sigmoid → prob: 53.0%
After 5pp cap: 53.0%
After [3,97] bounds: 53.0%

Reading Confidence Scores

75-100%
High Confidence

Strong evidence base, clear trend direction. We'd be surprised if this doesn't happen.

60-74%
Moderate Confidence

Good evidence but significant uncertainty remains. Could go either way with new developments.

40-59%
Low Confidence

Genuinely uncertain. Included because the question matters, not because we're confident.

+/-5/day
Max Daily Movement

Confidence scores are clamped to prevent overreaction to single-day signals. Big shifts accumulate over weeks.

Tech Stack

Python 3.12
Pipeline
Gemini 3.1 Pro
Primary LLM
Gemini 2.5 Flash
Fallback LLM
litellm
LLM Routing
Astro 6
Website
Tailwind 4
Styling
GitHub Actions
Daily Cron
GitHub Pages
Hosting

Open Source

The website frontend is open source. The forecasting pipeline methodology, predictions, and accuracy record are all published publicly so you can evaluate our work without needing the code.