ドル円機械学習トレーダー

A lot of capital in stocks is traded based on relatively simple factor models and formulaic alphas/regressions. Those factors are built by mining commercial datasets. Formulaic alphas are built though a mixture of experience, reinforcement learning and now more LLMs and genetic programming. Essentially figure out the market drivers, then aggregate a bunch of weak alphas that individually would not beat trading costs, but as an ensemble do perform well enough. This traditionally has been done at scale by hiring armies of quants and data scientists in slightly cheaper locations where the mathematic training is quite good, like Paris or Budapest. The alphas are then published by the quants and the portfolio managers pick them up in strategies. The quant are paid bonuses in relation to the orthogonal alpha they contributed to the PM's strategy performance. Most of the big names in the sector are running organizations based on those principles. WorldQuant, QRT, Cubist, you name it. I call it the "weak alpha ensembling" approach. This business model is about to be severly disrupted by AI-led reinforcement learning harnesses and will give an opportunity to "strong alpha stacking" approaches to make a comeback.

ニューラルネットは、少ないパラメーターでも、複雑なパターンが表現できちゃう。 金融予測では、あくまで共通項しか保持できないように、ニューロン数、パラメーターをギリギリまで絞るべし

1D CNNの追加で失敗で、実感するのは、やはり、パラメーター数が少しでも多いと、キャパ的にTrainデータの個別パターンを覚えられる。 未来で再現性あるパターン（シグナル）0.1% 再現しない個別パターン（ノイズ）99.9% ギリギリのパラメーター数で、あくまで共通項しか収まらないようにせねば

機械学習の結果、ドル円の値動きの形状は、予測に全く寄与しない結果に。 作成した状態特徴量は寄与、1D CNNでの形状特徴は寄与ゼロ つまり、短期中期チャートの形それ自体で判断しても、何ら意味なし。 Buy&Holdのようなベータ（長期ベースシフト）は取れるかもだけど。 ＃USDJPY　#ドル円

ビットコインは大量のエネルギーを消費した挙句、北朝鮮やイランや世界の反社団体に不正な送金手段を与えたくらいしか存在意義無かった そりゃAIに全部持ってかれても仕方ない