# MLP Architecture Screen — BTC 6H

## Morning conclusion

None of the LLM-inspired directions is promotable from the evidence available tonight. The current dense `55→16→8→1` model remains the clear baseline. The most promising research direction is a carefully validated causal regime-router, but the existing MVRV label is too volatile to use as a hard per-bar MoE gate and the simple entry-gating ablation materially reduced return. Synthetic augmentation is not justified as a way to manufacture new market regimes.

No winner CSV, live weights, Pine preset, or parameter file was changed.

| Idea | Closest implementation / test | Result | Decision |
|---|---|---|---|
| Dense model | Promoted `bb50` BTC 6H winner | Best available policy | Keep baseline |
| Sparse/selective routing | Logical grouped, random-sparse control, hybrid-grouped | All substantially below dense baseline | Reject |
| Sequence-aware proxy | Causal EMA companions of ten signals | Severe return/drawdown regressions | Reject |
| Regime hard routing | Existing causal MVRV label, fixed-policy entry-gate ablation | Every restricted gate lowers return | Do not build hard MoE yet |
| Synthetic data | Literature and data-generating-process review | No credible unseen-regime labels | Do not augment |

## Sparse expert routing is already a negative result

The grouped-input experiments are directly relevant to a tiny MoE: hidden units are allocated to specialist signal families instead of letting all features interact densely. The matched random-sparse control distinguishes a semantic routing benefit from simply removing parameters.

| Candidate | IS Calmar | IS P&L | Max DD |
|---|---:|---:|---:|
| Promoted dense baseline | 0.483 | 13,839.5% | -33.46% |
| Best logical grouped | 0.128 | 629.9% | -48.61% |
| Best random sparse | 0.167 | 2,243.5% | -60.24% |
| Best hybrid grouped | 0.158 | 2,550.3% | -66.30% |

The hybrid’s eight dense rows made it less bad than fully sparse routing, but it did not reliably beat the random-sparse control. This is strong evidence against spending compute on more LLM-style sparse input routing in this small, single-series setting. Sources: `docs/mlp_grouped_input_experiment_btc_6h_2026_08_20.md` and `docs/mlp_hybrid_grouped_input_experiment_btc_6h_2026_08_20.md`.

## Causal temporal companions are a negative sequence experiment

To test whether the tiny MLP was merely missing short/medium-horizon state, the experiment appended causal EMA(3) and EMA(12) summaries for ten existing signals. This adds no future information: every EMA at bar `t` uses only data at or before `t`. It is a small, Pine-exportable proxy for giving the network a short history rather than jumping straight to a transformer/attention model.

Replay used each artifact’s saved thresholds with the canonical simulator, `SCORE_START` through `TRAIN_END`. It is an IS rejection screen, not positive OOS evidence.

| Candidate | IS P&L | Max DD | Trades |
|---|---:|---:|---:|
| Current dense policy (fresh replay) | 13,829.8% | -24.47% | 56 |
| Temporal EMA seed 7070 | 462.0% | -63.80% | 108 |
| Temporal EMA seed 7171 | 748.5% | -83.24% | 74 |
| Temporal EMA seed 7272 | 442.2% | -77.90% | 211 |

The temporal models are not close enough to justify threshold sweeping or a Pine implementation. Their higher trade counts plus much worse drawdowns are a particularly poor failure mode for this strategy.

## Regime routing: causal label available, but not a stable router

The CSV already exports the MVRV regime used in Pine: bear `-1`, sideways `0`, bull `+1`. It is contemporaneous at every bar, so it is a viable causal input when the TradingView export is current; this is preferable to labelling regimes post hoc from future price movement.

However, across the 2017-12-01 to 2026-02-28 BTC 6H IS window, it switches 173 times across 174 contiguous segments. Median segment length is only 16 bars (four days), with some one-bar segments. A hard expert switch therefore creates a high-churn policy at exactly the points where per-expert score scales and thresholds can differ.

As a no-training gate ablation, I held the current winner’s weights and thresholds fixed and allowed entries only in the listed regimes; exits and position handling remained canonical. This intentionally gives the regime rule the easiest possible integration test before multiplying model parameters.

| Entry regime(s) | IS P&L | Max DD | Trades |
|---|---:|---:|---:|
| All (baseline) | 13,829.8% | -24.47% | 56 |
| Bull only | 2,873.1% | -20.37% | 50 |
| Bull + sideways | 7,968.3% | -27.28% | 55 |
| Bear + sideways | 370.0% | -16.25% | 10 |
| Sideways only | 163.1% | -21.89% | 7 |
| Bear only | 74.9% | -13.45% | 4 |

This does not prove every regime MoE fails, but it rejects the cheap deployable version: static MVRV entry suppression. Training three independent experts would cut effective supervised bars to 6,299 bull, 3,714 bear, and 2,031 sideways bars, while adding two models and a switching rule. The observed gate churn makes that trade-off unfavorable without a stronger router.

## Synthetic data is not a credible fix for data scarcity

Synthetic market paths can increase row count but cannot create independent examples of macro shocks, market microstructure changes, or unseen regime transitions. For this model it would also require generating all 55 normalized features jointly and causally; independently perturbing feature columns would destroy the cross-feature relationships the MLP uses. Training an expert on such data risks teaching it generator artifacts rather than price behavior.

This aligns with a recent broad time-series augmentation study: across its evaluated architectures and datasets, augmentation hurt in 67% of trials, and the authors report a strongly architecture-dependent effect. That is sufficient reason not to use synthetic data as an overnight shortcut here.

## Controlled next experiment, if revisited

Do not use post-hoc price labels. Instead, make the router a causal, slow-changing feature with explicit anti-churn rules: define a state vector from already-exported contemporaneous MVRV, realized volatility, trend, and liquidity signals; enforce fixed persistence/hysteresis; reserve complete chronological periods; and begin with one dense shared trunk plus tiny regime-specific output heads, not three complete MLPs. It should optimize one global strategy policy across switches and be Pine-exportable before promotion.

This is a genuine new architecture and should be an isolated, non-promoting experiment after a fresh TradingView export passes `validate_chart_data.py`.

## External grounding

- [Mixture-of-Experts with Expert Choice Routing](https://arxiv.org/abs/2202.09368) identifies expert under-training/load imbalance as a central MoE failure mode; the tiny, uneven regime slices here make that directly relevant.
- [Time-Series Forecasting for Out-of-Distribution Generalization](https://arxiv.org/abs/2406.09130) notes that time series lack adequate environment labels and that future shifts involve unobserved drivers—the danger in post-hoc market regimes.
- [Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters](https://arxiv.org/abs/2605.06032) reports highly architecture-dependent outcomes and mostly negative average augmentation results.