TL;DR
FinCast is a 1B‑parameter, decoder‑only Transformer trained on >20B financial time points with a token‑level sparse Mixture‑of‑Experts (MoE), learnable frequency embeddings, and a Point‑Quantile (PQ) loss that combines Huber point forecasts with quantile targets and a trend‑consistency term. In zero‑shot benchmarks across crypto/FX/stocks/futures, it reports ~20% lower MSE vs leading generic time‑series FMs, and it also beats supervised SOTAs—even without fine‑tuning—then widens the gap with a light fine‑tune. If you build risk or execution systems, the interesting part isn’t just accuracy points; it’s the shape of the predictions (tail‑aware, regime‑sensitive) and the deployment economics (conditional compute via sparse MoE + patching).
Why this matters (beyond leaderboard deltas)
Markets punish naïve stationarity assumptions. Most supervised forecasters regress to the mean when regimes shift, and many generic foundation models blur domain‑specific nuisances (microstructure noise, volatility clustering, calendar effects). FinCast’s bet is simple: capacity + conditional specialization + uncertainty‑aware training can travel across instruments and temporal resolutions without task‑specific surgery. For practitioners, that’s fewer bespoke models, more reuse, and forecasts that carry uncertainty information you can price and hedge against.
What’s new in FinCast (and why you should care)
1) Token‑level Sparse MoE (specialize without paying full price)
- What it is: Each token routes to the top‑k experts (k=2 out of 4 per MoE layer). That means higher representational capacity but only computing what you use.
- Why you care: In financial series, local patterns differ wildly (breakouts vs. mean‑reversion vs. volatility bursts). Specialist experts cut interference and help regime‑specific behavior emerge.
2) Learnable Frequency Embeddings (resolution is data, not metadata)
- What it is: An embedding encodes the sampling frequency (e.g., 1‑min, 1‑hr, 1‑day, 1‑wk) and shifts the model’s internal representation accordingly.
- Why you care: The same ticker at 1‑minute vs 1‑day behaves like different animals. Encoding frequency explicitly prevents the model from confusing high‑frequency noise with low‑frequency trend.
3) Point‑Quantile (PQ) Loss with Trend Consistency (predict the distribution, not just the mean)
- What it is: Combine a robust Huber point loss with quantile targets (e.g., 10th/50th/90th) and a trend‑consistency term on first differences; plus MoE regularizers to keep experts balanced.
- Why you care: Quantiles give you tail‑aware forecasts and avoid the notorious flat‑line failure mode. Trend consistency pushes directionality to match reality—useful for timing and sizing.
How it compares (finance‑specific choices vs. generic time‑series FMs)
| Feature | FinCast | TimesFM / Chronos‑T5 / TimesMoE |
|---|---|---|
| Architecture | Decoder‑only Transformer with token‑level sparse MoE | Decoder‑only; MoE in some variants but not finance‑tuned |
| Frequency handling | Learnable frequency embeddings (explicit) | Typically implicit or positional only |
| Training objective | PQ loss (Huber point + quantiles) + trend consistency + MoE balance | MSE/MAE style losses; limited distributional targets |
| Domain focus | Finance‑centric (crypto, FX, futures, stocks, macro) | General time‑series (finance included but not primary) |
| Zero‑shot stance | Built to generalize across domains & horizons | Strong generalization, but finance idiosyncrasies less targeted |
So what? If you manage a multi‑asset pipeline, FinCast’s knobs map more directly to portfolio questions (tail risk, horizon‑aware behavior, regime shifts), reducing the need for dataset‑specific hacks.
Evidence at a glance
- Scale: ~2.4M series, >20B points across crypto/FX/futures/stocks/macro plus non‑financial series for coverage.
- Compute: 1B params; MoE layers with 4 experts, top‑2 routing; trained ~147k steps with large batch; variable contexts up to 1024; patch tokenization.
- Zero‑shot: Average ~20% MSE reduction vs TimesFM/Chronos/TimesMoE across multi‑horizon tasks and multiple domains.
- Supervised: Even zero‑shot beats PCIE, PatchTST, Autoformer, Informer on US_71 & US_14L; a light fine‑tune (last 10% of MoE + output head, ~1 epoch) improves further.
- Latency: Sparse MoE + patching yields up to ~5× faster inference on an RTX 4060 vs peers at similar accuracy—practical for near‑real‑time risk dashboards.
Implications for builders (Cognaptus POV)
- Position sizing with quantiles. Treat the p10/p50/p90 spread as a confidence throttle: narrow spread → larger size; wide spread → smaller size or hedge.
- Horizon‑specific policies. With frequency embeddings, you can share one model but deploy different policy adapters by timeframe (e.g., scalping vs. swing vs. weekly rebalance).
- Regime routing as telemetry. MoE gate activations are a free signal: which experts fire across assets today? That’s a regime map you can log and watch for rotation.
- Guardrails against flat‑lining. If your current forecaster collapses to the mean under stress, swapping in a PQ‑trained head (even on your existing model) is low‑hanging fruit.
- Cost‑aware scaling. Sparse compute means you can explore broader universes (more tickers/timeframes) without doubling GPU bills—a big deal for small desks.
Where it fits (and where it doesn’t)
- Great for: Cross‑asset forecasting backbones, risk scenario generators, signal pre‑filters before RL/policy layers, stress testing with quantile bands.
- Use with caution for: Ultra‑low‑latency HFT (still autoregressive + patch iterations), thinly traded microcaps (data quality dominates), and anything where causality matters more than pattern extrapolation.
A practical rollout plan
- Shadow deployment: Run FinCast alongside your incumbent model for 2–4 weeks, logging p10/p50/p90 and realized errors.
- Decision overlays: Replace hard thresholds (e.g., fixed stop widths) with quantile‑sized bands; cap exposure when quantile dispersion spikes.
- Expert analytics: Track expert usage by asset/timeframe; promote experts that align with profitable regimes; down‑weight chronic laggards.
- Minimal fine‑tune: If needed, unfreeze the output block + last MoE slice on your asset universe for a single epoch—cheap and usually enough.
Open questions we’ll keep testing
- Leakage vs. generalization: How robust are zero‑shot gains when benchmarks evolve and ex‑ante leakage checks get stricter?
- Live drift: Does the quantile calibration stay stable month‑on‑month without re‑tuning?
- Expert health: Are there failure patterns when one expert becomes over‑dominant, and can router regularizers catch it early in prod?
Bottom line
FinCast’s design choices are finance‑native and pragmatic: encode frequency explicitly, route computation to specialists, and train to a distribution. If you’re consolidating a zoo of one‑off models, this is a credible path to a single backbone with policy‑layer adapters on top. We’ll prototype FinCast‑style PQ heads and expert‑telemetry dashboards in Cognaptus pipelines over the next cycle and report what survives contact with live markets.
Cognaptus: Automate the Present, Incubate the Future