从新手到专家：使用 MQL5 制作动画新闻标题（四） 本地托管 AI 模型市场洞察·进阶篇
给 News Headline EA 接上 AI 滚动通道
把外部 AI 洞察塞进 MT5 的 EA，核心不是算法本身，而是用最小改动让 EA 能定时拉取、安全绘制、干净卸载。下面这套写法在保留原 EA 逻辑的前提下，只新增输入项、全局态、一个抓取函数和三处生命周期钩子。 先开三个输入：布尔开关控制通道启停，字符串填 FastAPI 或任意 HTTP 端点，整数定两次 POST 的最小间隔秒数。实测把 InpAIInsightsReloadSec 设成 60，代表每分钟最多打一次远端，避免图表卡在等待响应。 全局里用 latestAIInsight 存文本，offAI 管横向滚动偏移，aiRequestInProgress 拦并发请求，lastAIInsightTime 记上次成功时刻。这四个变量保证断网或报错时画面不空——旧洞察继续滚。 FetchAIInsights() 把所有 HTTP 细节封死在函数内：开关没开、上一次还在飞、冷却未到都直接 return。通过才拼 JSON 发 POST，成功就从响应抠 "insight" 字段，失败留旧值。下面代码逐行拆了关键段。 OnInit 里建一块半透明 AI 画布垫在原通道下方，先写占位符再立刻调一次 FetchAIInsights()，所以哪怕中途加载 EA，首次网络回来就有真内容。OnTimer 每次重绘完顺手清画布、按偏移绘文本、左移循环。OnDeinit 杀定时器、删画布，重加载必从零起。外汇与贵金属波动剧烈，接入外部信号仅作辅助参考，实盘前请在策略测试器跑通 WebRequest 权限。
class=class="str">"cmt">//--- class="num">1) USER INPUTS ------------------------------------------------ input class="type">bool ShowAIInsights = true; input class="type">class="kw">string InpAIInsightsURL = "http:class=class="str">"cmt">//class="num">127.0.class="num">0.1:class="num">8000/insights"; input class="type">int InpAIInsightsReloadSec = class="num">60; class=class="str">"cmt">// seconds between requests class=class="str">"cmt">//--- class="num">3) GLOBALS ----------------------------------------------------- class="type">class="kw">string latestAIInsight = "AI insights coming soon…"; class="type">int offAI; class=class="str">"cmt">// scroll offset class="type">bool aiRequestInProgress = false; class=class="str">"cmt">// prevent concurrent POSTs class="type">class="kw">datetime lastAIInsightTime = class="num">0; class=class="str">"cmt">// last successful fetch time class="type">void FetchAIInsights() { if(!ShowAIInsights || aiRequestInProgress) class="kw">return; class=class="str">"cmt">// 开关关或请求在飞则退出 class="type">class="kw">datetime now = TimeTradeServer(); class=class="str">"cmt">// 取交易服务器时间 if(now < lastAIInsightTime + InpAIInsightsReloadSec) class="kw">return; class=class="str">"cmt">// 冷却未过退出 aiRequestInProgress = true; class=class="str">"cmt">// 置标志防并发 class="type">class="kw">string hdrs = "Content-Type: application/json\r\n"; class="type">class="kw">string body = "{\"prompt:\"Concise trading insight for " + Symbol() + "\"}"; class=class="str">"cmt">// 拼当前品种提示 class="type">uchar req[], resp[]; class="type">class="kw">string hdr; StringToCharArray(body, req); class=class="str">"cmt">// 转字节数组 class="type">int res = WebRequest("POST", InpAIInsightsURL, hdrs, class="num">5000, req, resp, hdr); class=class="str">"cmt">// 发请求超时5秒 if(res > class="num">0) { class="type">class="kw">string js = CharArrayToString(resp,class="num">0,WHOLE_ARRAY); class=class="str">"cmt">// 响应转字符串 class="type">int p = StringFind(js, "\"insight\":"); class=class="str">"cmt">// 找insight键 if(p >= class="num">0) { class="type">int start = StringFind(js, "\"", p+class="num">10) + class="num">1; class=class="str">"cmt">// 值起点 class="type">int end = StringFind(js, "\"", start); class=class="str">"cmt">// 值终点 if(start>class="num">0 && end>start) latestAIInsight = StringSubstr(js, start, end-start); class=class="str">"cmt">// 截出洞察文本 } lastAIInsightTime = now; class=class="str">"cmt">// 更新成功时间 } aiRequestInProgress = false; class=class="str">"cmt">// 清标志 } class="type">int OnInit() { class=class="str">"cmt">// … existing init … class=class="str">"cmt">// AI Insights lane if(ShowAIInsights) {
「AI 洞察跑马灯的定时器与生命周期」
在 OnInit 末尾用 EventSetMillisecondTimer(InpTimerMs) 挂起毫秒级定时器，AI 洞察条才具备持续刷新的动力；初始化时先 CreateBitmapLabel 建名为 AiC 的位图标签，TransparentLevelSet(120) 把整体透明度压到约 47%（120/255），避免遮挡主图价格行为。 OnTimer 里每 tick 先 FetchAIInsights() 拉最新文本，若 ShowAIInsights 为真则 Erase(ARGB(120,0,0,0)) 清屏并重绘：offAI 按 InpAIInsightsSpeed 左移，当 offAI + TextWidth(latestAIInsight) < -20 时复位到 canvW，形成无缝循环跑马灯。 OnDeinit 中必须 EventKillTimer() 并 aiCanvas.Destroy() 加 ObjectDelete(0,"AiC") 清理，否则 MT5 切周期会残留不可见对象。外汇与贵金属行情受突发数据影响大，这类叠加层仅作信息辅助，实际下单仍应以价格结构为准，相关风险偏高。
aiCanvas.CreateBitmapLabel("AiC", class="num">0, class="num">0, canvW, lineH, COLOR_FORMAT_ARGB_RAW); aiCanvas.TransparentLevelSet(class="num">120); offAI = canvW; SetCanvas("AiC", InpPositionTop, InpTopOffset + (InpSeparateLanes ? class="num">8 : class="num">5) * lineH); aiCanvas.TextOut(offAI, (lineH - aiCanvas.TextHeight(latestAIInsight)) / class="num">2, latestAIInsight, XRGB(class="num">180,class="num">220,class="num">255), ALIGN_LEFT); aiCanvas.Update(true); class=class="str">"cmt">// initial fetch FetchAIInsights(); } EventSetMillisecondTimer(InpTimerMs); class="kw">return INIT_SUCCEEDED; } class="type">void OnTimer() { class=class="str">"cmt">// … existing redraw for events/news/indicators … class=class="str">"cmt">// fetch & draw AI lane FetchAIInsights(); if(ShowAIInsights) { aiCanvas.Erase(ARGB(class="num">120,class="num">0,class="num">0,class="num">0)); aiCanvas.TextOut(offAI, (lineH - aiCanvas.TextHeight(latestAIInsight)) / class="num">2, latestAIInsight, XRGB(class="num">180,class="num">220,class="num">255), ALIGN_LEFT); offAI -= InpAIInsightsSpeed; if(offAI + aiCanvas.TextWidth(latestAIInsight) < -class="num">20) offAI = canvW; aiCanvas.Update(true); } } class="type">void OnDeinit(const class="type">int reason) { EventKillTimer(); class=class="str">"cmt">// … existing cleanup … if(ShowAIInsights) { aiCanvas.Destroy(); ObjectDelete(class="num">0, "AiC"); } }
◍ 把本地推理延迟和 UI 卡顿拆开看
EA 接上本地 FastAPI-Uvicorn 后，新开的 AI Insights 通道会实时显示模型回传的文本。测下来一次完整推理周期从发 POST 到收回 JSON，端到端耗时约 31.98 秒，外汇与贵金属交易本身高风险，这种秒级延迟只适合非逐 tick 决策的辅助场景。 llama-cpp 加载 GGUF 是一次性成本，日志里 load time = 206235.75 ms（约 206 秒），之后每个 POST 才进入常规推理。提示评估 4 个 token 用了 1487.17 ms，平均每 token 371.79 ms；生成 63 个 token 的 eval time = 29555.55 ms，每 token 约 469.14 ms（~2.13 token/s）；采样 64 次仅 58.01 ms。加起来 total time = 31979.70 ms，服务器回 200 OK，且出现 prefix-match hit 说明缓存命中跳过了重复层计算。 滚动停顿的根因在 EA 定时器里直接调 FetchAIInsights() 会阻塞 WebRequest，最长卡约 32 秒等服务器返回。解法是把绘制文本和 HTTP 调用解耦：每 20 ms 先滚动画面，推理完成后再用响应更新 latestAIInsight，通道动画就不再断。 别把 32 秒当固定值 第二次循环 total time 升到 35799.69 ms，prompt eval 也变慢，说明本地 CPU 推理波动明显。真要上 MT5 验证，先确认 AI 服务器后台常驻，否则 EA 通道会一直空等超时。
<span style="background-class="type">color:rgb(class="num">249, class="num">204, class="num">202);">llama_print_timings: load time = <span class="number">class="num">206235.75</span> ms llama_print_timings: sample time = <span class="number">class="num">58.01</span> ms / <span class="number">class="num">64</span> runs ( <span class="number">class="num">0.91</span> ms per token, <span class="number">class="num">1103.33</span> tokens per second) llama_print_timings: prompt <span class="built_in">eval</span> time = <span class="number">class="num">1487.17</span> ms / <span class="number">class="num">4</span> tokens( <span class="number">class="num">371.79</span> ms per token, <span class="number">class="num">2.69</span> tokens per second) llama_print_timings: <span class="built_in">eval</span> time = <span class="number">class="num">29555.55</span> ms / <span class="number">class="num">63</span> runs ( <span class="number">class="num">469.14</span> ms per token, <span class="number">class="num">2.13</span> tokens per second) llama_print_timings: total time = <span class="number">class="num">31979.70</span> ms ←[<span class="number">class="num">32</span>mINFO←[<span class="number">class="num">0</span>m: <span class="number">class="num">127.0</span><span class="number">.class="num">0</span><span class="number">.class="num">1</span>:<span class="number">class="num">52770</span> - <span class="class="type">class="kw">string">"←[1mPOST /insights HTTP/class="num">1.1←[0m"</span> ←[<span class="number">class="num">32</span>m200 OK←[<span class="number">class="num">0</span>m Llama.generate: prefix-<span class="keyword">match</span> hit</span> llama_print_timings: load time = <span class="number">class="num">206235.75</span> ms llama_print_timings: sample time = <span class="number">class="num">83.42</span> ms / <span class="number">class="num">64</span> runs ( <span class="number">class="num">1.30</span> ms per token, <span class="number">class="num">767.19</span> tokens per second) llama_print_timings: prompt <span class="built_in">eval</span> time = <span class="number">class="num">1890.97</span> ms / <span class="number">class="num">6</span> tokens( <span class="number">class="num">315.16</span> ms per token, <span class="number">class="num">3.17</span> tokens per second) llama_print_timings: <span class="built_in">eval</span> time = <span class="number">class="num">32868.44</span> ms / <span class="number">class="num">63</span> runs ( <span class="number">class="num">521.72</span> ms per token, <span class="number">class="num">1.92</span> tokens per second) llama_print_timings: total time = <span class="number">class="num">35799.69</span> ms ←[<span class="number">class="num">32</span>mINFO←[<span class="number">class="num">0</span>m: <span class="number">class="num">127.0</span><span class="number">.class="num">0</span><span class="number">.class="num">1</span>:<span class="number">class="num">52769</span> - <span class="class="type">class="kw">string">"←[1mPOST /insights HTTP/class="num">1.1←[0m"</span> ←[<span class="number">class="num">32</span>m200 OK←[<span class="number">class="num">0</span>m Llama.generate: prefix-<span class="keyword">match</span> hit llama_print_timings: load time = <span class="number">class="num">206235.75</span> ms llama_print_timings: sample time = <span class="number">class="num">51.40</span> ms / <span class="number">class="num">64</span> runs ( <span class="number">class="num">0.80</span> ms per token, <span class="number">class="num">1245.21</span> tokens per second) llama_print_timings: prompt <span class="built_in">eval</span> time = <span class="number">class="num">1546.64</span> ms / <span class="number">class="num">4</span> tokens( <span class="number">class="num">386.66</span> ms per token, <span class="number">class="num">2.59</span> tokens per second)
本地推理日志里的耗时分布
把 llama.cpp 跑在本地给 MT5 做行情洞察接口时，日志里最能说明瓶颈的是 eval time 与 load time。三次连续请求里 eval time 分别是 29878.89 ms、31295.30 ms、29311.62 ms，每次都是 63 个 token 生成，折算每 token 约 465~497 ms，生成速度仅 2.01~2.15 tokens/s。 load time 三次都固定在 206235.75 ms，说明模型只加载一次后常驻，后续请求不再重复载入，这部分对单次延迟没有叠加影响。prompt eval time 随输入 token 数在 1487~1841 ms 间波动，4~6 个 token 的提示词就要 300~415 ms/token，短 prompt 的预处理并不轻。 sample time 反而极快，三次分别为 65.92 ms、55.34 ms、58.01 ms 对应 64 runs，单 token 采样不到 1.1 ms，吞吐高达 970~1156 tokens/s，证明解码采样不是制约点。prefix-match hit 反复出现，意味着相似系统提示被 KV 缓存命中，能省掉部分 prompt 重算。 真要压低 MT5 端等待，重点应砍 eval 阶段的层数或量化等级，而不是去优化采样。外汇与贵金属波动快，这类本地 LLM 辅助判断仅作概率参考，实盘仍属高风险。