KJj W xdZddlZddlZddlZddlZddlZejjejjejjejje dddlmZmZmZddlmZgdZdZgdZgd Zd Zd Z dd ZiZd Z ddZdZdZdZ y)u MLP-based activation score strategy. Replaces the single-layer weighted sum of strategy_activation_scores.py with a small multi-layer perceptron over the pre-normalised feature columns (55 for BTC, 50 for alts; see FEATURE_COLS / the artifact's feature_cols): score = 1000 * tanh( W3 . tanh( W2 . tanh(W1.x + b1) + b2 ) + b3 ) Everything downstream of the score — crossunder entry/exit, confirmation, trailing stop, regime/MVRV gates, metrics — is identical to the existing strategy (imported or mirrored verbatim), so results stay directly comparable and the TV-parity workflow carries over. Weights are NOT optimizer params: they live in a JSON artifact (strategies/params/mlp/mlp_weights_{ASSET}_{TF}.json) produced by tools/train_mlp.py and referenced via the `mlp_weights_file` param. The same weights are embedded into strategy_mlp_scores.pine as array literals by tools/generate_pine_mlp_presets.py, with tanh implemented in Pine by the identical clamped closed form (Pine v6 has no math.tanh): tanh(x) = x >= 20 ? 1.0 : x <= -20 ? -1.0 : (exp(2x) - 1) / (exp(2x) + 1) Pilot limitations (documented intentionally): - `i_div_window` is unsupported — the MLP consumes raw single-bar RSID flags, matching _prepare_features()'s window=1 behaviour. Nz..)_apply_trailing_stop_prepare_featurescalculate_metrics)calculate_positions) FEATURE_COLSFEATURE_WEIGHT_PARAMSACTIVATION_NAME_tanh mlp_forwardsave_mlp_artifactload_mlp_artifactgenerate_signalsrtanh_closed_form_clamp20)7 stoch_normmacd_pred_normosc_norm totalvol_normm3_momentum_norm m2_tiny_norm newaddr_normstoch_div_norm vwap_div_normstoch_peak_norm sendaddr_norm m3_div_normbearish_engulfing_score m2_nooff_norm m2_div_normbullish_hammer_scorebullish_engulfing_scoreshooting_star_scorebtc_spx_corr_30 dxy_roc_normvix_pctrank_invbtc_dom_roc_signus10y_roc_inv_signspy_above_200emagold_roc_pctrankmvrv_zscore_valuemvrv_zscore_cont nupl_normfed_net_liq_sign gc_positionus2y_roc_inv_signyield_curve_sign sopr_normrsid_reg_bull_normrsid_reg_bear_normrsid_hid_bull_normrsid_hid_bear_normrsid_rt_bull_normrsid_rt_bear_normrsid_slow_bull_normrsid_slow_bear_normrsid_delayed_peak_normrsid_delayed_dip_norm oi_roc_norm usdt_d_norm basis_normcvd_norm btc_gold_normrsi_subtf_norm bb_pct_b_normr> rvol_norm active1y_normhr_normr0)2 i_w_stoch i_w_macd_predi_w_osc i_w_totalvoli_w_m3_momentum i_w_m2_tiny i_w_newaddri_w_stoch_div_osci_w_vwap_div_osci_w_stoch_peaking i_w_sendaddri_w_m3_div_osci_w_bearish_engulfingi_w_m2_div_osc_noOffseti_w_m2_div_osci_w_bullish_hammeri_w_bullish_engulfingi_w_shooting_stari_w_btc_spx_corri_w_dxyi_w_vix i_w_btc_dom i_w_us10yi_w_spyi_w_goldi_w_mvrv i_w_mvrv_conti_w_nupli_w_fed_net_liqi_w_gc_positioni_w_us2yi_w_yield_curvei_w_qqq_spy_ratioi_w_rsid_reg_bulli_w_rsid_reg_beari_w_rsid_hid_bulli_w_rsid_hid_beari_w_rsid_rt_bulli_w_rsid_rt_beari_w_rsid_slow_bulli_w_rsid_slow_beari_w_rsid_delayed_peaki_w_rsid_delayed_dip i_w_oi_roc i_w_usdt_d i_w_basisi_w_fear_greed i_w_btc_gold i_w_rsi_subtf i_w_bb_pct_bctj|tj}tj|}|dk\}|dk}||z}d||<d||<tjd||z}|dz |dzz ||<|S)a Clamped closed-form tanh, formula-identical to the Pine implementation. Inside (-20, 20): (exp(2x) - 1) / (exp(2x) + 1). At |x| >= 20 returns exactly +/-1.0 (Pine and Python clamp at the same point so neither side can diverge in the saturated region). dtypeg4@g4?gg@)npasarrayfloat64 empty_likeexp)xouthilomide2xs w/Users/jameslopez/projects/TradingBot25/.worktrees/mlp-results-and-research-learnings/strategies/strategy_mlp_scores.pyr r s 1BJJ'A -- C dB eB G*CCGCG &&qv Cc cCi(CH Jctj|tj}|D]\}}tjddd5|tj|tjjztj|tjz}dddt }|dddfdzS#1swY"xYw)a Forward pass: X (T, F) float64, layers = [(W, b), ...] with W (n_out, n_in). Every layer (including the output layer) applies _tanh; the final scalar is scaled by 1000 so scores live in [-1000, +1000] like the existing activation-score convention. rxignore)divideoverinvalidNrg@@)r{r|r}errstateTr )XlayersaWbzs rr r s 1BJJ'A1 [[x JBJJq 3555 1BJJ8WWAK !H QT7V K Js ACC c\| t|ntt}ttj|ddj dg|Dcgc]0\}}ttj|j d2c}}z}d|||t ||D cgc]l\}} tj|tjjtj| tjjdnc} }|xsid} tjtjjtjj|dt|d 5} tj | | dddycc}}wcc} }w#1swYyxYw) zWrites the weights artifact JSON. Floats serialised via repr (full precision) by json, so load() round-trips bit-exact float64.Nrrx)rr)versionasset timeframearch activation feature_colsrtrainingT)exist_okw)listrintr{r|shaper r}tolistosmakedirspathdirnameabspathopenjsondump) rrrrrrr_rrpayloadfs rr r sV*6)A4 %tLGYL  6!9Q<(..q1 2 3067061BJJqM   "#7 D%$ 1**Qbjj188:**Qbjj188: < N GKK 56F dCA '1 %7  s5FA1F6F""F+c tjj|}tjj|}tj |}| |d|k(r|dSt |5}tj|}ddddDcgc]R}tj|dtjtj|dtjfT}}||d|d |j d |j d |j d |j d |j didd}||ft|<|S#1swYxYwcc}w)aLoads (and caches by path+mtime) a weights artifact. Returns {"layers": [(W, b), ...] float64, "arch": [...], "feature_cols": [...], "meta": {...}}. Note: artifacts trained on the 49-feature set (pre-bb_pct_b_norm) are still loadable via load_mlp_artifact() and mlp_results_table; they carry their own feature_cols list and will fall back to 0 for any missing column. Nrrrrrxrrrrrrrr)rrrrr)rrrmeta) rrrgetmtime_ARTIFACT_CACHEgetrrloadr{arrayr}) rrmtimecachedrrlayerrarts rr r sAggood#G GG  W %E   )F fQi50ay g!))A, X&&E %*BJJ /%*BJJ1WX&  /{{9-[[) [1!++l3 J3  C!&s|OG J' s.EAE$E!c t|} tj| tj} tj| t} tj| t} d} d}d}d}t | D]}|dn t ||}| rB||k7r |||kDr||}|d|z z}|||k}|r||n||}|s|r0|r.d} d| |<d}d}d}n ||r|rd} d| |<||}|dz}|||k}| rdnd| |<| | | fS) uLike _apply_trailing_stop but with dead-zone guard for exit smoothing. When a trade enters while score_exit[entry_bar] < exit_threshold (i.e. the SMA is already below the exit threshold — the "dead-zone"), the smoothed crossunder can never fire for that trade because there is no transition from above to below the threshold. In that case we switch to the raw-score crossunder (exit_raw_fallback_arr) for the duration of that trade only. time_ok_arr (optional bool array): see _apply_trailing_stop — Pine gates strategy.close on timeCondition, so entries/exits only fire where time_ok_arr[t] is True; a position open at window end is held open. rxFTrzrrlenr{zerosint32boolrange) entry_raw_arr exit_raw_arrexit_raw_fallback_arrscore_exit_arrexit_threshold close_arrhigh_arrtrail_stop_pct time_ok_arrnin_pos exec_entry exec_exitposition trade_highfill_baruse_rawtallowed stop_price trail_hitexit_sigs r_apply_trailing_stop_with_dzrs? IA!288,F!4(J!4(IHJHG 1X%-$4 A3G H}!z!9%a[ #s^';Q.bhh?J   a N 2   a > 1 3 - - / 1_[[^~ - [[^n , . ( ( *#+   a N 2   a > 1 3   1_[[^~ - [[^n , .  #+    q !_ 4   q !O 3 5 U[[^ # %^^A%^ 0 0 1 !&&q)_<A^_ I%I./Jzz*+H K("6FbjjI??5)0077=H$++E299@@FL&&NQ C ,H8\8^' - )FJ -I8\- )FJ  # (< xHn#) % I )"H<)>LqAX,=L>&@h!<*?1h@ .5!LW8K!W !L!  ~-Y)*, MML>)RSZR[\88    MML>)9#h-I,,/ ,=+>geWMMM   B%%'++113BJ"h'A(CM:Bq!tWN7 ""))"**5I6 !!((4H RH MMI?@ "Ms$ G G G"G2GG)NNNN)N)!__doc__rrsysnumpyr{pandasrrappendrjoinr__file__%strategies.strategy_activation_scoresrrr$strategies.library_activation_scoresr__all__r rrr r r rr rrrrrrr;s:  RWW__X-F MNO E -8 z*& &JN#'6#Z043)l):}@8Nr