The Journey

The project opens with the most natural question in financial ML: can a sequence model predict the next move? R000 wires up the full pipeline — LSTM 2x128, lookback 60, binary_12 labels asking 'will price be higher in 1 hour?' — and gets answered with a brutal -36.5% PnL and a 45.2% win rate. Binary labels are noise. R001 swaps to triple-barrier labeling (TP=0.75%, SL=0.5%) and uncovers a deeper set of issues: the model was shorting on label=0 (which doesn't mean 'down', only 'not up'), class weights were missing, and the Sharpe formula was double-counting flat bars. R002 patches those, R003 simplifies to a 21K-parameter GRU to fight overfit. None of it works at fixed thresholds.

The pivot happens at R004: thresholds should scale with volatility. The new formula TP = k_up · σ, SL = k_dn · σ becomes the labeling backbone that survives the entire project. Twenty-three rounds (R004-R026) sweep k_up/k_dn from 1.8/1.2 to 25/15, alpha from 6 to 96, lookbacks from 20 to 120. R012 establishes GRU > LSTM as a permanent fact. R020-R022 prove multi-timeframe features add +795%. R025 validates walk-forward 4-fold. By R026 the project thinks it has a +998% champion (R022).

Then R027 happens. Investigating a discrepancy between training and production, the team discovers that filtering-out 'timeout' bars during training created sequence gaps that did not exist in live data. Every result in R004-R026 was inflated — sometimes by 60x. The honest +998% becomes +15.9%. Twenty rounds of work are reduced to a single durable insight (volatility-adaptive barriers) and a brutal lesson about pipeline parity.

With the sequence-gap bug exorcised, the team has to rebuild credibility against honest numbers. R027-R033 try swing labels and a dual-head architecture; R033 hits +17.6% with the dual head. R034 attempts pure regression and finds an unusable combination — WR 75% and PF 3.56, but only 16 trades across the whole walk-forward. R035 explores 1-minute base bars and immediately blows up GPU memory.

R036 introduces what becomes the second foundational idea: predict MFE (maximum favorable excursion) and MAE (maximum adverse excursion) as continuous values. DualHeadGRU + MSE on both heads gives +8.3%; R036b tunes TP/SL to +9.4%; R036c shows the model knows when NOT to enter, losing only -1.2% in a bear market where B&H loses -23%. The 'knows when to sit out' result is the project's first hint that the model has real, asymmetric edge.

R038 then performs the single most important loss-function experiment. Standard MSE produces uniform predictions (MFE/MAE ratio ≈ 1.0) — the model cheats by predicting the mean. Four losses are compared: standard MSE (-4.9%), mse_ratio (+16.2%), weighted (-13.6%, hacks loss), and asymmetric (-0.6%). mse_ratio is the ONLY viable loss. R037 then validates mse_ratio + MFE/MAE in proper walk-forward across 4 folds (2018-2026) and posts +97.9% total / +9.3% worst fold / WR 54% / PF 1.49.

The lurking problem after R037 was sample efficiency: 95% of 5-minute BTC bars carry essentially no information, but the GRU was being asked to learn from all of them equally. R046 introduces the most important architectural reframe: train only on 'interesting' bars, defined as bars where ATR > 2x the rolling mean OR |returns| > 90th percentile. About 10.5% of bars survive the vol_tight filter.

The results are dramatic. The model now sees a sequence of regime transitions rather than a flood of mostly-flat candles. R046-R053 sweep event filters, lookbacks, and feature sets. By R053 the best config — binary labels + event-driven sequences + lookback=100 + vol_ultra filter — posts +685% summed across 4 walk-forward folds. That is roughly 7x the previous walk-forward record and ~1.7x B&H. 15min is confirmed as the optimal base timeframe.

This epoch is short (8 rounds, 3 days) but represents the single highest-leverage methodological change in the project. Every subsequent breakthrough is built on top of event-driven sequences.

With event-driven sequences proven, R054-R060 settle every component decision that will define the V2-V6 production stack. R054/054c/054d run three independent normalization studies and unanimously confirm full per-sequence MinMax — R054c posts +791% compound at lb=300. R055 tests 3-class labels (SL=-1, timeout=0, TP=+1) and wins +824% vs +685% binary; the key is CrossEntropyLoss with ce_signal weights [2.0, 0.5, 2.0].

R056 runs permutation importance + correlation + ablation studies on the original 47 features, identifying 11 actively harmful features (volatility_60, volume_ratio, hour_sin, MACD, ema crosses, etc.) — the 'clean 36' feature set is born. R057 confirms clean 36 is more robust than positive_only 24. R058 introduces the compound metric (product of (1+fold)) and posts +4,783% compound. R059 combines 3-class + clean 36 features. R060 does the definitive event-filter grid and confirms vol_tight + k25/15 + thr=0.40 as the sweet spot at +4,021% compound.

This epoch is the project's least glamorous and arguably its most important. None of these rounds 'discovered' anything; they consolidated and stress-tested the discoveries of the previous epoch.

R061-R067 systematically refine the trading layer. R061 compares fixed/dynamic/trailing/hybrid/always exits and finds fixed simplest and best (+3,440%). R062 confirms multi-timeframe features don't help once lb=300 is in place. R063 retires shorts (P(SL) isn't precise enough) and SL-signal exit. R064 settles expanding-window over sliding. R065 confirms 15min over every alternative and discovers that sigma clip [0.0005, 0.005] is essential — removing it cuts compound by 2.5x. R066 picks 100% position sizing. R067 finds that requiring 3-4 consecutive event-bars in agreement improves trade quality +47%.

Then R068 changes everything. Rather than asking 'when should the model enter?', it asks 'why is the model selectively trading at all?' Answer: stay invested like B&H by default, use the model only to detect dangerous regimes and exit. This single reframing captures all of B&H's upside while letting the model add asymmetric protection. It is the single biggest conceptual breakthrough in the project.

R069 introduces 5-seed-per-config validation (single-seed results revealed as unreliable). R070 confirms voting (≥4/5 agree) beats simple averaging. R071 then assembles the full stack and posts +7,128% compound, beating B&H in 4/4 folds. R072 validates across three seed sets: set A passes 4/4, set B 3/4, set C 4/4.

After R071 set a record but R072 revealed instability, the natural hypothesis was 'the labels can be smarter.' R073-R080 is the project's most concentrated burst of experimentation: thirteen different labeling schemes tested in a single calendar day. Speed-weighted labels, 5-class labels, efficiency-weighted, DSR (deflated Sharpe), filtered labels, next-event labels, RL meta-decision, Self-Distillation Iterative (SDIL), Conditional Barrier Asymmetry (CBA), Multi-Resolution Consensus (MRC), Trend Scanning, Path-Quality Weighted (PQW).

The result: none of the thirteen methods beat plain 3-class tb3_vol. Two insights survive. First, regression labels fundamentally fail with the GRU — CrossEntropy concentrates gradients on the directional decision in a way no continuous loss can match. Second, timeouts (~24% of labels) are not noise but essential information; every 'cleaner' labeling scheme that suppresses or downweights timeouts loses more signal than it gains.

This epoch is the canonical example of a productive dead-end. The team spent eight rounds learning that the foundation was already as good as it could be, which freed the next epoch to focus on production deployment.

R081 combines every breakthrough into a single 'ultimate production config'. R082 begins fine-grid threshold tuning. Then R083 — the second great bug-fix moment of the project.

In the always-invested notebooks (R068 onward), the equity computation at the end of each fold force-closed the open position and then computed equity[-1] = capital + position_value — double-counting the last trade. The reported +89,000% compound was actually +4,000-7,600%. Two months of 'records' had to be re-evaluated. The corrected R084 result is +7,641% best-set / +3,978% worst-set / B&H baseline +1,005% — still a 4-7.6x outperformance.

R085-R087 explore whether anything obvious can lift the corrected numbers. R085/R085b add sigma, tp_dist_pct, sl_dist_pct as features — all hurt (-24% to -45%). R086/R087 sweep regularization and model size — all hurt or fail to improve cross-set consistency. R089-R092 try LSTM swap, multi-lookback, 9-seed ensembles — none beat baseline.

Then R094 breaks the plateau in an unexpected direction. Trying SMA as a sanity baseline, the team discovers that simple SMA on its own crushes the GRU — and combining either with a DAILY SMA filter explodes performance: GRU+daily gives +720M%, SMA+daily gives +9.3B%. Every tested config with the daily filter beats B&H in 4/4 folds. The daily SMA filter is the single most important component the project has ever found.

With the GRU stack validated and the daily SMA filter discovered, R095-R111 are the project's deployment epoch. R095 tests ATR-adaptive exits — improves V4 by +53% but hurts V3/V6. R096 confirms multi-timeframe entry hurts. R097 confirms drawdown circuit breakers hurt. R098 confirms RSI/MACD/volume filters layered on top of SMA all hurt. The pattern is consistent: the daily SMA filter is essential, but additional filters add noise.

R099-R109 build out three production bots: V2 (pure GRU ensemble, +7,641% backtest), V3 (GRU + DailyRSI>80 + 10% trailing stop, +450B% backtest), V4 (3-regime adaptive V5.4 Robust-5, +30,174% backtest, +117% min α). All three deployed to Hetzner.

R110-R111 are the climax. An external collaboration produces V115_cmp (combines GRU + peak_drop + ratchet + regime cooldown) and V66 (uniform thresholds + extreme cooldown 4,48 bars). R110 reproduces them at the dollar: V115_cmp at +168,759% compound / +238% min α; V66 at +70,576% compound / +245% min α (the all-time min-α record). R111 builds the HybridV115Trader class (~750 LOC) and V5+V6 join the live stack. Five servers, one shared GRU checkpoint, five strategy configurations.

With V5/V6 deployed and posting numbers nobody dared to budget for, the natural next question was: can we extract more from the same GRU checkpoints? R112-R116 attack the V66 strategy layer from every angle. R114 implements confidence-weighted voting. R115 sweeps vote thresholds and min_votes parameters. R116 implements Kelly position sizing.

The result of five rounds is unanimous: V66's hyperparameters are at a local optimum that cannot be improved by any tweak inside the same parameter family. Confidence-weighted voting matches binary voting in compound but doesn't improve robustness. Vote-threshold sweeps confirm the existing config sits at the peak. Kelly sizing reduces drawdown but also reduces compound.

This is the second major dead-end of the project — and like R073-R080 before it, the value is in the negative result. The team formally closes the V66-refinement search and reframes the question.

R117-R120 mark the project's most recent strategic pivot. Having proven that the V66 stack cannot be improved within its own vocabulary, the team formalizes a new objective in ACTION_PLAN_2026.md: produce a candidate that beats V5/V6 by a meaningful margin AND is diversifying — uncorrelated enough with the existing stack to be deployed alongside it.

The four exploratory rounds attack four orthogonal hypotheses. R117 tests mean reversion (BB_revert, RSI_extreme, Z_score, MACD_div) — the dual of the trend-following stack. R118 implements XGBoost as an alternative model class — REJECTED, GRU crushes trees 287×. R119 implements a multi-timeframe regime filter — REJECTED, V66 already encodes context. R120 is the portfolio-composition breakthrough: regime-aware allocation of V66 + MACD_div delivers 7 winners with **362,991% compound (5.1× V66) and matching 245% min α**.

Alongside the experiments, the project hardens its validation infrastructure (Phase 0: 5-step gate). The era of 'a single notebook + a single seed' is formally over. The current strategic focus is concluding R120 analysis and building a real-time regime detector (R121) so the retroactive R120 alpha can be deployed.