Culture is a distributed sentiment engine. We want to tap the wires.
Hollywood greenlights a movie two years before you watch it. Spotify knows what you want to hear tonight. Each sensor has its own latency. Read them in the right order and you might have a credit-cycle signal.
The idea
Credit cycles are driven by collective risk appetite. Risk appetite is a psychological variable. We measure it indirectly through credit spreads, volatility, survey data, and flow indicators — all of which are already priced by the time they show up. We think there’s an earlier read available in what humans make and consume when nobody’s asking them to reveal a risk preference.
This scenario is a bet that culture is a distributed sentiment engine — a network of sensors that measure collective uncertainty, optimism, fear, and risk tolerance at different latencies. If you read the sensors in the right order, you might be able to estimate the risk-appetite variable directly, earlier and more cleanly than a credit-spread-based proxy.
Two signal layers
This is the part that makes the bet interesting. Cultural production is not one signal — it’s two, and they move at different speeds.
Production signal (lead, 12–36 months). What gets greenlit, financed, announced, put into development. A movie that comes out this year was a bet placed two to three years ago by a studio whose cost-of-capital and future-revenue assumptions were frozen at that moment. A novel published next spring was acquired 18 months ago. A restaurant opening in Q3 was financed last winter. Production decisions are dense with forward-looking sentiment — the people making them are effectively writing directional options on the 2028 mood.
Consumption signal (coincident, near real-time). What gets watched, reviewed, charted, liked, streamed. Streaming and social are now fast enough that we can measure demand-side mood at weekly or even daily resolution. This is not forward-looking; it’s a thermometer for what people want to feel right now.
Any serious framework has to separate these two layers. Mixing them destroys the lead-lag structure that makes the bet worth placing.
The sensors
Each content domain is a different sensor. Different latency, different noise profile, different mechanism.
| Sensor | Signal type | Latency | What it captures |
|---|---|---|---|
| Film/TV development slates | Production | 12–36 months | Industry expectations, capital availability, risk tolerance in greenlights |
| Novel acquisitions (agent/publisher announcements) | Production | 12–24 months | Editorial bets on what readers will want |
| Music charts (Spotify/Apple) | Consumption | Daily–weekly | Immediate mood, risk appetite, affect |
| IMDb / Metacritic scores, demographic splits | Consumption | Daily | Audience response, cohort divergence |
| Fashion (runway seasons, color forecasts) | Mixed | 6–18 months | Risk tolerance in visual expression |
| Food trends (reservation data, menu sentiment) | Consumption | Weeks | Comfort vs. experimentation, consumer confidence |
| Bestseller lists (non-fiction especially) | Consumption | Weekly | Structural anxieties, self-help cycles |
Music is the fastest thermometer. Film slates are the slowest bet. Books sit in between, with the specific property that non-fiction bestseller composition tracks structural worries remarkably well — self-help and financial-resilience titles cluster before and during credit contractions.
Why this might work
Everything on that list is downstream of the same latent variable: collective perception of future stability and willingness to take risk. That’s exactly what drives credit expansion and contraction. If there’s a shared hidden state, and we have enough independent measurements of it, we should be able to recover an estimate of the state that’s cleaner than any single measurement — which is the standard argument for sensor fusion.
The specific edge we’re betting on is the production/consumption split. Credit markets already price consumption-side sentiment reasonably well. Credit markets do not price production-side sentiment at all, because the horizons don’t match — a credit trader is looking at the next quarter, and a greenlight decision is a three-year forward option. If the production side is at all informative about credit conditions 18 months out, it is almost certainly not priced in.
The test we’d run
- Pick one clean production-side dataset as the primary instrument. Best candidate: major-studio film greenlights and development-slate additions, timestamped by the date the project was announced rather than the date it released. We want the moment of commitment, not the moment of delivery.
- Extract features from it at that commitment moment: genre distribution, budget percentile, tone classification (dark/optimistic/nostalgic/dystopian — LLM one-shot classification is a defensible approach here), franchise-vs-original ratio, and average cast/director risk premium.
- Build a “production-side risk appetite index” as a time series of those features aggregated at monthly resolution.
- Cross-correlate it against the ~788-day credit cycle we already isolated in the Spectral work. The specific question: does the production-side index lead the credit cycle by 12–24 months, and by how much?
- Pre-register the lag window and the test statistic before looking at the answer. No p-hacking over lag choices.
If the production index has a coherent cross-correlation peak at a plausible lead, that’s interesting. If it’s noise, we learn something and move on.
Consumption-side sensors (music charts, review sentiment) are a separate experiment — they go against shorter-horizon volatility and credit spreads, not the 2-year cycle. We’d build them second.
How it could die
- The signal is there but reflexive rather than causal. Studios respond to macro sentiment by greenlighting darker content — but the sentiment they’re responding to is the same thing the credit market is responding to, with different latencies. In this case the production-side signal isn’t a leading indicator of the credit cycle, it’s a lagging indicator of whatever made the credit cycle move. That would still be measurable, but it wouldn’t be tradeable.
- Confounded by industry structure. The shift to streaming dramatically changed the economics of greenlighting. Any secular trend in risk appetite in the production data is potentially confounded by the streaming-era content glut. We’d need a null model that controls for total production volume, not just composition.
- Feature extraction is the whole game and we under-invest. Classifying tone from a logline is a hard problem. If the encoder is noisy, no downstream analysis can recover.
- The null is hard. Unlike daily market data, we have maybe 40–50 years of reliable industry data and much less with consistent metadata. 30 annual observations is not enough for the 2-year cycle to resolve cleanly. We’d be leaning hard on monthly resolution.
- Mechanism isn’t stable. “Sell in May” worked until it didn’t. Any reflexive cultural signal is vulnerable to becoming known and arbitraged away. We’d only trust it if it had held across at least two full credit cycles out of sample.
What would make us take it seriously
A production-side index that cross-correlates with the ~788-day credit cycle at a consistent lead of 12–24 months, surviving the same gate progression we use everywhere else: Bonferroni over a pre-registered lag grid, block-shuffle permutation for temporal clustering, and out-of-sample replication on an international equivalent (UK publishing, Japanese media) where the cultural vocabulary differs but the underlying variable should be the same.
Why this fits z-of-a
We already have a credit cycle that explains ~8% of volatility variance and a validation pipeline that’s comfortable killing its own hypotheses. Adding a sensor fusion layer on the input side is the natural extension — and it belongs squarely in the zone of avoidance, because no quant fund is going to hire a postdoc to classify movie loglines by tone. The data is messy, the methodology is unglamorous, and the reputational cost of being “the culture guy” is about the same as being “the astrology guy” was nine months ago. We don’t mind.