The Barometer

The Scoreboard Doesn't Lie

Sixteen days in. Twelve trading days. Twenty open positions. The portfolio is up $29 on $100K — basically flat. But the composition of that number is screaming something useful.

Here are the winners:

And the losers:

See the pattern?

Three of our top four performers were driven by forward-looking sources — prediction markets, economic nowcasts, structural game theory. Our biggest loser, the SPY short, was driven by reactive headline scanning. RSS saw 36 military conflict articles and screamed "short equities." By the time 36 articles exist, the market already priced it in.

The defense longs (XAR, ITA) are the exception — RSS caught those early because defense is a second-order play that the broader market was slow to rotate into. But for the primary indices? Reading yesterday's news is yesterday's edge.

The Reactive Stack

Let's be honest about what we built. Here are Tauntaun's eight signal sources, classified by what they actually measure:

Six reactive. Three predictive. And the three predictive sources are producing the majority of our alpha.

What "Predictive" Actually Means

Let's not get mystical about this. Prediction isn't prophecy. It's reading the barometer instead of the weather report.

The weather report says: "It rained today." The barometer says: "Pressure is dropping. Rain is likely tomorrow."

In market terms:

Reactive: "36 articles about military conflict published in the last 8 hours" → SHORT SPY. The market already knows. You're late.

Predictive: "Kalshi has a 44% probability of aggressive Fed cuts below 3.0%" → LONG TLT. The crowd is pricing in something that hasn't happened yet. You're early.

Both are data-driven. Both are quantifiable. But one is reading the exhaust trail and the other is reading the engine.

The Reactive Sources Aren't Useless

Before I burn the newsroom down — the reactive stack has a job. It's just not the job I originally thought.

RSS and GDELT are excellent at one thing: detecting when the world's attention shifts. Not predicting what happens next, but measuring what the world is talking about right now. That's a sentiment indicator, not a directional signal.

The GPR Index is the same — it measures the current level of geopolitical tension based on newspaper coverage. It's useful for portfolio-level risk management ("are we in crisis territory?") but lousy at timing individual trades.

Credit spreads measure institutional complacency. When they're tight, the market is calm. When they blow out, something broke. They don't predict the break — they confirm it after the fact.

FRED data is released on a known schedule. By the time unemployment hits the wire, every algo on Wall Street has already parsed it. The edge isn't in the number — it's in what the number means for the next quarter.

So the reactive stack becomes what it should've been all along: context, not catalyst. It tells the strategy engine what the world looks like right now. The predictive stack tells it where the world might be going.

The Shift

Here's what changes.

1. Reweight the fusion engine. Right now, all sources get equal footing in signal fusion. A 0.85 confidence from RSS (36 articles about a thing that already happened) carries the same weight as a 0.60 from Kalshi (real money wagered on a thing that hasn't happened yet). That's wrong. Forward-looking sources should carry a premium in the fusion math. Not 10x — but enough that a strong prediction market signal can override a loud-but-late news cycle.

2. Reclassify reactive sources as filters, not generators. RSS, GDELT, GPR, and Credit Spreads stop generating primary trade signals. Instead, they become confirmation layers. The predictive stack proposes trades; the reactive stack validates them. "Does the current news environment support this bet?" is a different question than "What should we buy based on today's headlines?"

3. Expand the predictive stack. Three sources isn't enough. Candidates:

4. Build a "conviction score" that weights freshness. A prediction made today is worth more than one made two weeks ago — but a prediction made two weeks ago that's still holding is worth more than both. The current exponential decay is too simple. We need a model where predictions gain conviction when the market moves in their direction and lose it when the market moves against them. A living scorecard, not a countdown timer.

What This Looks Like in Practice

Today's signal stream is a good example. Here's what came in at 07:00 UTC:

Predictive stack:

Kalshi says 44% probability of aggressive Fed cuts → LONG TLT, SHORT SPY.
ORALE nowcasts CPI annualized at 10.3% → LONG TIP.
Predictive History says ceasefire collapse resumes oil disruption → LONG USO, defense.
Stock scanner flags BA at 66/100 with strong momentum → LONG BA.

Reactive stack:

RSS sees 36 military conflict articles → LONG ITA, LONG USO, SHORT SPY.
GPR Index elevated at 304 (z=1.7) → LONG GLD, LONG TLT.
Credit spreads complacent at 2.94% → light GLD hedge.

Under the old model, every one of those generates a signal at face value. The RSS 0.85 on ITA outweighs the Kalshi 0.52 on TLT. Headlines dominate.

Under the new model: Kalshi's 44% Fed cut probability is the primary directional signal. ORALE's CPI nowcast confirms inflation is hot. The professor's game theory adds structural context. RSS's 36 articles validate that the geopolitical environment supports the defense thesis — but the trade was already on the table before the first headline was written.

The reactive stack's role becomes: "Is now a bad time to enter this trade?" not "Should we make this trade?"

The Risk

One obvious problem: prediction markets can be wrong. Kalshi had a 36% probability of hot inflation — that means 64% of the money says it won't be hot. Betting on the 36% side because "it's forward-looking" is still a minority bet. The Professor's game theory scenarios are structural — they play out over months, not days — and the market can stay irrational longer than your paper trading account stays patient.

This is why the reactive stack becomes a filter, not a reject pile. The predictive sources propose. The reactive sources sanity-check. And the fusion engine requires convergence — multiple predictive sources agreeing, confirmed by the current environment — before a high-confidence signal fires.

We're not throwing away six data sources. We're promoting three and demoting six. The chain of command changes. The team stays the same.

When We'll Know If It Works

We need 30 more trading days to have enough closed positions to measure this properly. The plan:

Week 1-2: Implement the reweight. Tag every signal with its source class (reactive vs. predictive). Keep the old fusion logic running in parallel as a shadow book — same signals, old weights — so we can A/B compare.

Week 3-4: Add at least one new predictive source (options flow is the most accessible via free data). Monitor whether the shadow book diverges meaningfully.

Week 5-6: First real comparison. Did the predictive-weighted portfolio outperform the equally-weighted shadow? By how much? On which trades?

If predictive weighting doesn't outperform, we'll know. The data will say so. And we'll write that entry too.

The weather report tells you to bring an umbrella. The barometer tells you to cancel the picnic. One saves your shirt. The other saves your day.